Coding-model intelligence · Updated 2026-08-14

The honest map of what LLM coding actually costs.

Metered per token or bought as a subscription — price, provider precision and published quality, aggregated, not re-benchmarked. Everything below is sourced and dated, including the column no other comparison shows: what numeric precision you are actually buying.

Best value · idx per $

1Ling-3.0-flash1606 idx/$

2Solar Pro 41004 idx/$

3gpt-oss-120b468 idx/$

Best value · top 20% (idx ≥ 69)

1DeepSeek V4 Flash$0.18/M

2GPT-5.6 Luna$0.22/M

3DeepSeek V4 Pro$0.54/M

Cheapest / session

1Xiaomi MiMo-V2.5$0.072

2DeepSeek V4 Flash$0.139

3Tencent Hy3$0.295

Quality ceiling · coding idx

1GPT-5.6 Sol78.3

2Claude Opus 578.0

3Grok 4.676.8

Cheapest open-weight · avg

1CoreWeave · 10 models$0.71/M

2DeepInfra · 38 models$0.71/M

3StreamLake · 14 models$0.74/M

Best plan deal · tok/$ @ $0.82/M

1opencode Go7M/$

2GitHub Max2M/$

3Cursor Ultra2M/$

Editor’s picks

Hand-written · 8 on record

Written from the data alone; referral relationships carry no weight in what is picked or how it is described. Each pick keeps the date it was made and is not quietly rewritten later, so you can judge how well it aged.

2026-08-14

Free today, or a GPU you have to buy

The free hosted stack is the move at $0 — Nemotron 3.5 Lightning free on Zen (34.7% on our BCB-Hard) · the local dense 27B needs a bigger card · Grok 4.6 near the top

What you can do for free this week keeps getting better, and for most of us that is still the move. Nemotron 3.5 Lightning, a 30B MoE with only 3B active, landed free on opencode Zen and we measured it ourselves at 34.7% on BigCodeBench-Hard, right in the band of the cheap open coders, and it reasons before it answers. Pair it with OpenRouter's :free routes and the rest of the rotated Zen set — Hy3, DeepSeek V4 Flash (coding index 69), MiMo-V2.5 — and the heavy, high-volume passes cost nothing. That is a real coding budget at $0. The other story is local, and it is more honest than exciting: Qwen3.8-27B dropped, a dense 27B, and with MTP speculative decoding through llama-server it finally runs interactively on a 16 GB card (about 36 tok/s). But to fit 16 GB you run it at Q3, and at Q3 it scores 27% on our BCB-Hard — below what already runs there — so the genuine local upgrade needs a bigger card that fits Q4 or Q6, and that is a pick for people with money to spend on a GPU. Everyone else on a 16 GB card waits for the small A3B MoE Qwen will ship next, which runs fast on the hardware you already own; the hands-on note is in Run local. Above all of it the ceiling barely moved: Grok 4.6 joined at 76.8, just under GPT-5.6 Sol (78.3) and Claude Opus 5 (78.0), still best on a subscription than per token. And Synthetic's $7 first-month promo is over, back to $30. As always the free sets rotate without notice, so treat this as the week's allocation.

2026-08-06

Two ways to code cheap this week — one free, one nearly so

opencode Zen + OpenRouter :free for the volume · Synthetic $7 for Kimi K3 · frontier still climbing

Two ways to code cheap this week, one free and one nearly so. The free stack is the strongest it has been: opencode Zen now serves eight models at no cost — DeepSeek V4 Flash (coding index 69), MiMo-V2.5, Nemotron 3 Ultra, Ling-3.0-flash and the newly-added LongCat-2.0 among them — and OpenRouter mirrors much of it as :free routes, including Gemma 4 31B (which we measured at 43% on our own BCB-Hard) and gpt-oss-20b. Lean on those for the heavy, high-volume passes; they cost nothing right now. For the calls that have to land, the nearly-free option is Synthetic's weekend promo — $7 for the first month buys a month of Kimi K3 (coding index 72) plus GLM-5.2 and Qwen 3.6 on a private, OpenAI-compatible API, more frontier coding than $7 has any right to. Above that the ceiling keeps rising: Muse Spark 1.2 (72.2) and Qwen3.8 Max (71.8) joined the board this week alongside Kimi K3, and Claude Opus 5 (78.0) is still the highest tracked — sensible on a subscription rather than per token. As always, free sets rotate and close without notice (opencode Zen swaps models weekly) and the $7 renews at $30, so treat this as the week's allocation, not a standing plan.

2026-07-25

The free stack got stronger — and the paid step up

GLM-5.2 on top · MiMo-V2.5 & DeepSeek V4 Flash free · Opus 5 to pay

The free window that carried last week's pick has already closed: Tencent Hy3 is no longer free on OpenRouter — GMICloud now charges $0.13/$0.53 — exactly the risk that pick named. The free tier did not shrink, though; it got stronger. opencode Zen now serves Xiaomi MiMo-V2.5, DeepSeek V4 Flash, Ling-3.0-flash, North Mini Code and Nemotron 3 Ultra at no cost, and OpenRouter mirrors most as :free variants. Lean on MiMo-V2.5 (56.8 coding index, the strongest measured model in the free set) and DeepSeek V4 Flash (56.2, reasoning) for the heavy, high-volume work — it costs nothing right now, and even paid they run at about $0.06 and $0.08 a session. Keep one genuinely strong model on top for the calls that have to be right: GLM-5.2, at a 68.8 coding index with a 1M context and roughly $0.8/$2.6 per million, is the sensible flagship on opencode Go or OpenRouter. And to buy the ceiling without metered-token pain, Claude Opus 5 is a solid step up — 78.0 coding, the highest tracked — kept affordable by taking it on the Claude subscription rather than per token. Free tiers rotate and close without notice, as Hy3 just showed, so treat this as the week's allocation, not a standing plan.

2026-07-18

This week's stack, and when to change it

Hy3 and MiMo-V2.5 now · GLM-5.2 before October

Two things are free right now that are not obviously worse than what you pay for. Tencent Hy3 scores 58.8 on coding index — fractionally above MiniMax-M3 — and is served at no cost with a 262K context on both OpenRouter (tencent/hy3:free) and opencode (hy3-free). Pair it with MiMo-V2.5 at $0.058 per observed session and most routine agent work is close to free. When those free windows close, GLM-5.2 is the obvious next stop at $0.29/$0.92 with a 1M context — but move before October, for the reason in the next pick.

Correction 2026-07-25: The Hy3 free window has since closed — it is no longer free on OpenRouter (GMICloud now charges $0.13/$0.53), exactly the risk this pick flagged. MiMo-V2.5 stays free on opencode Zen; see the 2026-07-25 pick for the current free stack.

2026-07-18

Two price rises that are not price rises

GLM-5.2 in September · MiniMax already

Z.ai's own FAQ says GLM-5.2 and GLM-5-Turbo "will only consume 1x quota during off-peak hours, valid through the end of September" — so 30 September 2026. From October the standard rates return: 2x off-peak, 3x peak. The sticker price does not move; $18, $72 and $160 all stay put. But peak is only four hours a day (14:00-18:00 UTC+8), so the promotion currently covers twenty hours in every twenty-four, and when it lapses those twenty hours double their burn. Your plan halves while the invoice stays identical. Worth knowing what is being multiplied, too: the marketing counts prompts, but the account API reports both windows as TOKENS_LIMIT, so the multiplier lands on a token meter. MiniMax has already run its own version of this manoeuvre — cached tokens used not to count against the subscription quota, at some point they began to, and the effective cost rose sharply with no announcement. Neither change appears on any pricing page, which is exactly why a subscription's real cost has to be measured rather than read.

2026-07-18

Cheap does not mean cheap per session

GLM-5.1

At $1.27 per observed session it is roughly 22x MiMo-V2.5 and the most expensive model tracked here, despite a list price well under Kimi K2.6's. The driver is cache: 86.2% against about 95% for everything else, so far more input is billed at full rate. List price and running cost rank these models in different orders.

2026-07-18

Check what precision you are buying

Provider quantization

The same model is served at different numeric precision for near-identical money. MiniMax-M3 costs $0.30/$1.20 across nine providers, six of them fp8, one bf16, two undisclosed. On Qwen3 Coder Next, Parasail serves bf16 at $0.12 while Alibaba, the model's own vendor, charges $0.30 without stating precision at all. Cheapest is often most quantized; sometimes it is not.

2026-07-18

The ceiling, and what open models cost you

The frontier vs the open field

The top coding score tracked here is GPT-5.6 Sol at 78.3, with Claude Fable 5 just behind at 76.5 and 62.9% on TerminalBench-Hard. The best open model reaches 58.8 and 42.4%. That gap is real and large — roughly 18-20 points on both. It also costs $10/$50 per million against $0.14/$0.28 for MiMo-V2.5, so the frontier is priced at two orders of magnitude more per token. Which is worth it depends entirely on whether the task fails silently: for supervised edits the cheap models are close enough, for long autonomous runs it compounds.

Correction 2026-07-20: as first published this pick called Claude Fable 5 the ceiling. It was not: GPT-5.6 Sol scored higher on the same index at the time. The fault was ours — the model roster was filtered through a hand-written list of labs that omitted OpenAI, Google and ten others, so 39 measured models were missing from the site. The roster is now derived from the data with no such list. The frontier-vs-open argument above is unaffected.

1 / 8

Weekly notes

betathe run, in plain words

A short write-up each week on what changed and why — the same data as the tables, told as a story. Newest first.

2026-08-14

First weekly note: a free model that reasons, and prices that moved

I am starting a short note each week to explain what changed on the site, in plain words. Same data as the tables, just told as a story. This is the first one.

A new free model I could actually test. NVIDIA released Nemotron 3.5 Lightning this week, a 30B mixture-of-experts model with only 3B active parameters. The small active count is the interesting part: it means the model is cheap to run and it fits on modest hardware. It is also already free on opencode Zen, so I did not need to rent a GPU. I pointed our BigCodeBench-Hard harness at the free endpoint and measured it myself: 34.7% on BCB-Hard (42 of 121). It is a reasoning model, so it thinks before it answers and it is not fast, but for a free local-agent model that is a solid number. It is now both on the leaderboard (Artificial Analysis gives it a coding index of 26.8) and in the Run local catalogue with our own measurement.

Qwen3.8-27B landed, and I ran it. The Max is still datacenter-only (2.4T, 95B active), but the small dense 27B (Apache-2.0) that fits at home shipped this afternoon, so I tested it properly. On my 16 GB card it fits at Q3 (about 14 GB), and with MTP speculative decoding through llama-server it runs at 36 tok/s, twice the 18 it manages without spec-decode, because the model ships its own draft head and about 87% of the drafts get accepted. On our BCB-Hard it scores 27% at Q3 non-thinking; the full model at good precision (BF16/FP8) is around 37 to 43%, so the 3-bit quant you need to fit 16 GB costs about 10 points. It is not on the leaderboard yet, no Artificial Analysis index and not on OpenRouter, so for now it lives in Run local with our own numbers. It made me want a bigger GPU, but on the 16 GB card I actually have it does not beat what I already run, so I have not switched, the honest take is in the Run local pick. And it is dense, so if you cannot buy a card the one to wait for is the small A3B MoE Qwen will ship next, which runs fast on the hardware you already own.

The ceiling kept climbing. Grok 4.6 joined the board at a 76.8 coding index, near the top of everything tracked. Solar Pro 4 (52.7), Muse Glimmer 30B (49) and Nemotron 3.5 Lightning also went on. Qwen3.8-Max was already there at 71.8. DeepSeek also shipped a new revision of V4 Pro (0813) that jumped its coding index from 59.4 to 68.8, a big step for an already cheap open model.

GLM-5.3 dropped today too. Zhipu launched it on August 14, "built to code", and it looks strong: they claim +50% coding over GLM-5.2 and first place among open-source models on Terminal-Bench 3.0 (DeepSWE 1.1 66.9, CyberGym 84.5). The Z.ai coding plan already serves it, and GLM-5.2 and 5.1 requests now route to 5.3. But it is not on OpenRouter or the Artificial Analysis index yet, and the open weights are about two weeks out, so it does not go on the leaderboard yet. That is the rule here: a model only joins the board once it has a public coding index and a place to buy it. I will add it the moment it lands on both.

Prices moved in a few places. OpenAI turned on unlimited text with GPT-5.6 Luna for the Free and Go tiers, it went live this week. Note it is text only and the paid tiers still keep caps on Luna, the earlier claim that Luna was unlimited on Plus and Pro was wrong. Synthetic's $7 first-month promo ended, it is back to $30 a month. Mistral folded its old $5.99 Education tier into a student discount on the $14.99 Pro plan. Alibaba now shows a $2 new-user discount across its Individual token tiers.

DeepSeek is about to get more expensive. V4 Pro 0813 shipped at the old flat rate, but from August 16 (16:00 UTC) DeepSeek switches to peak/off-peak billing and the jumps are steep: cache-hit input rises up to 12x at peak, cache-miss up to 3x, output up to 4.5x, with peak on Beijing working hours (01:00-04:00 and 06:00-10:00 UTC). The cache-hit line is the one that stings, because a coding agent replays a big cached prefix every turn, so the effective bill tracks that 12x, not the headline flat price. If you lean on DeepSeek, front-load the heavy runs before the 16th, or push them to off-peak after.

Two corrections worth calling out. The "Cerebras free tier sunsets on Aug 17" note that went around is a false alarm: that date only kills preview models on the paid Developer tier, the free trial is not going away. And Morph quietly rebuilt its whole pricing to pure pay-per-token with 200 free requests a month, the old Free/Starter/Pro/Scale credit tiers are gone. Their page bot-walls plain requests, so I had to open it in a real browser to read it, now confirmed.

New this week. The free set on opencode Zen rotated: it added Hy3 and Nemotron 3.5 Lightning and dropped Ling-3.0-flash, LongCat-2.0 and North Mini Code. I added RunPod to the gateways, it has a real OpenAI-compatible inference API through its Public Endpoints, though the text catalogue is thin and it is mostly image and video. And I added this notes tab.

As always, free sets rotate and close without notice, so treat any free allowance as this week's, not a standing plan.

And one last thing: next week we bring a surprise. Stay tuned.

03 QualityLeaderboard Published quality Off the index

Leaderboard

Ranked by coding index. Value = coding index per $ (blended, 3:1 in:out).

#	Model	Coding index	Intelligence	TBench-H	Long-ctx	tok/s	$/M blended	Value idx/$	Context
1	GPT-5.6 Sol · OpenAI	78.3	59.0	0.614	0.763	57.0	$11.250	7	1,050,000
2	Claude Opus 5 · Anthropic	78.0	63.1	—	0.757	45.1	$10.000	8	1,000,000
3	Grok 4.6 · xAI	76.8	60.9	—	0.750	52.1	$3.000	26	500,000
4	GPT-5.6 Terra · OpenAI	76.7	56.6	0.576	0.797	89.8	$2.250	34	1,050,000
5	Claude Fable 5 · Anthropic	76.5	62.1	0.629	0.767	56.2	$20.000	4	1,000,000
6	Kimi K3 · Moonshot	76.2	59.7	—	0.827	36.5	$6.000	13	1,048,576
7	GPT-5.5 · OpenAI	74.9	56.3	0.606	0.790	0.0	$11.250	7	1,050,000
8	Claude Opus 4.8 · Anthropic	74.3	57.3	0.583	0.730	0.0	$10.000	7	1,000,000
9	Claude Opus 4.7 · Anthropic	73.6	55.0	0.515	0.753	0.0	$10.000	7	1,000,000
10	Grok 4.5 · xAI	72.4	55.8	—	0.740	50.9	$3.000	24	500,000
11	Muse Spark 1.2 · Meta	72.2	56.8	—	0.833	0.0	$2.000	36	1,048,576
12	Qwen3.8 Max · Alibaba / Qwen	71.8	58.1	—	0.743	43.3	$3.000	24	1,000,000
13	Claude Sonnet 5 · Anthropic	71.5	55.3	—	0.770	72.4	$4.000	18	1,000,000
14	GPT-5.6 Luna · OpenAI	71.4	52.3	—	0.783	139.0	$0.225	317	1,050,000
15	Muse Spark 1.1 · Meta	71.3	53.2	—	0.813	0.0	$2.000	36	1,048,576
16	GPT-5.4 · OpenAI	71.1	53.1	0.576	0.777	0.0	$5.625	13	1,050,000
17	Gemini 3.5 Flash · Google	70.1	52.0	0.409	0.810	0.0	$3.375	21	1,048,576
18	Gemini 3.6 Flash · Google	69.2	51.6	—	0.790	199.7	$1.500	46	1,048,576
19	DeepSeek V4 Flash · DeepSeek	69.1	51.8	—	0.743	101.1	$0.175	395	1,048,576
20	GLM-5.2 · Z.ai / Zhipu	68.8	52.6	0.508	0.767	104.3	$1.828	38	1,048,576
21	Gemini 3.1 Pro Preview · Google	68.8	47.7	0.538	0.790	128.4	$4.500	15	1,048,576
22	DeepSeek V4 Pro · DeepSeek	68.8	53.2	—	0.753	66.6	$0.544	127	1,048,576
23	Qwen3.7 Max · Alibaba / Qwen	66.0	46.7	0.508	0.747	0.0	$2.213	30	1,000,000
24	Claude Sonnet 4.6 · Anthropic	63.0	48.4	0.530	0.740	0.0	$6.000	10	1,000,000
25	Kimi K2.6 · Moonshot	61.8	45.1	0.439	0.767	0.0	$1.010	61	262,144
26	Kimi K2.7 Code · Moonshot	60.8	43.0	0.447	0.750	36.8	$1.407	43	262,144
27	Xiaomi MiMo-V2.5-Pro · Xiaomi	60.2	42.9	0.432	0.777	47.7	$0.544	111	1,050,000
28	KAT-Coder-Pro V2 · Kwaipilot	59.5	34.5	0.492	0.700	109.7	$0.525	113	262,144
29	Nex-N2-Pro · Nex AGI	59.1	41.7	—	0.763	142.2	$0.438	135	262,144
30	Tencent Hy3 · Tencent	58.8	42.2	—	0.747	66.5	$0.231	255	262,144
31	MiniMax-M3 · MiniMax	58.6	45.4	0.424	0.803	70.0	$0.525	112	1,048,576
32	Xiaomi MiMo-V2.5 · Xiaomi	56.8	38.0	0.417	0.683	90.0	$0.175	325	1,050,000
33	GPT-5.4 Nano · OpenAI	56.1	39.7	0.424	0.720	0.0	$0.463	121	400,000
34	GPT-5.4 Mini · OpenAI	56.1	40.9	0.523	0.730	0.0	$1.688	33	400,000
35	Qwen3.7 Plus · Alibaba / Qwen	55.9	39.4	0.470	0.690	56.3	$0.560	100	1,000,000
36	GLM-5.1 · Z.ai / Zhipu	55.8	41.0	0.432	0.680	0.0	$2.150	26	204,800
37	Qwen3.6 Plus · Alibaba / Qwen	54.5	40.5	0.439	0.723	0.0	$0.731	75	1,000,000
38	Qwen3.6 27B · Alibaba / Qwen	53.7	37.7	0.348	0.733	53.3	$1.350	40	262,144
39	Solar Pro 4 · Upstage	52.7	41.6	—	0.707	42.0	$0.052	1004	524,288
40	MiniMax M2.7 · MiniMax	52.6	38.9	0.394	0.753	0.0	$0.525	100	204,800
41	Inkling · Thinking Machines	52.1	42.3	—	0.733	81.8	$1.725	30	1,048,576
42	Grok Build 0.1 · xAI	51.5	40.7	—	0.700	0.0	$1.250	41	256,000
43	Ling-3.0-flash · InclusionAI	50.6	37.8	—	0.670	380.4	$0.032	1606	262,144
44	GPT-5.1 · OpenAI	49.4	37.5	0.455	0.767	0.0	$3.438	14	400,000
45	Gemini 3.5 Flash-Lite · Google	49.3	37.4	—	0.747	305.6	$0.850	58	1,048,576
46	Nemotron 3 Ultra 550B · Nvidia	49.3	38.3	0.364	0.710	132.0	$1.350	37	512,288
47	Muse Glimmer · Meta	49.0	35.1	—	0.800	90.2	$0.637	77	131,072
48	Qwen3.5 397B A17B · Alibaba / Qwen	48.2	34.3	0.409	0.727	80.3	$0.877	55	262,144
49	Mistral Medium 3.5 · Mistral	46.9	30.4	0.333	0.653	122.2	$3.000	16	262,144
50	Kimi K2.5 · Moonshot	46.8	36.0	0.348	0.730	0.0	$1.140	41	262,144
51	GLM 4.6 · Z.ai / Zhipu	45.8	29.3	0.250	0.553	0.0	$0.963	48	204,800
52	Qwen3.5-122B-A10B · Alibaba / Qwen	45.7	32.8	0.311	0.703	132.8	$0.817	56	262,144
53	LongCat 2.0 · Meituan	45.3	34.0	—	0.627	35.4	$0.525	86	1,048,756
54	GLM 4.7 · Z.ai / Zhipu	45.3	34.5	0.318	0.680	0.0	$0.738	61	204,800
55	DeepSeek V3.2 · DeepSeek	44.2	32.8	0.356	0.707	0.0	$0.302	146	163,840
56	DeepSeek V3.1 Terminus · DeepSeek	43.5	31.1	0.303	0.680	0.0	$0.440	99	163,840
57	Gemma 4 31B · Google	43.4	29.7	0.364	0.683	35.7	$0.160	271	262,144
58	Ring-2.6-1T · InclusionAI	42.8	31.7	0.288	0.673	121.3	$0.212	201	262,144
59	Grok 4.3 · xAI	42.2	37.9	0.379	0.660	0.0	$1.562	27	1,000,000
60	Qwen3.6 35B A3B · Alibaba / Qwen	41.9	32.1	0.348	0.667	138.3	$0.362	116	262,144
61	o1 · OpenAI	39.7	23.9	0.129	0.633	0.0	$26.250	2	200,000
62	Step 3.7 Flash · StepFun	39.6	30.9	0.356	0.697	357.3	$0.438	91	262,144
63	Gemma 4 26B A4B · Google	39.3	26.1	0.136	0.617	0.0	$0.190	207	262,144
64	GPT-5 · OpenAI	37.8	35.3	0.326	0.763	0.0	$3.438	11	400,000
65	Qwen3.5-35B-A3B · Alibaba / Qwen	37.0	24.3	0.106	0.603	157.1	$0.619	60	262,144
66	North Mini Code · Cohere	36.5	20.2	0.311	0.360	26.5	$0.000	—	256,000
67	Qwen3 Coder Next · Alibaba / Qwen	36.2	21.3	0.182	0.423	134.3	$0.290	125	262,144
68	Gemini 3.1 Flash Lite · Google	34.7	25.6	0.242	0.713	0.0	$0.562	62	1,048,576
69	Gemini 2.5 Pro · Google	33.3	25.9	0.265	0.660	0.0	$3.438	10	1,048,576
70	Devstral 2 · Mistral	31.3	19.2	0.189	0.313	44.4	—	—	—
71	Mercury 2 · Inception	31.1	21.9	0.265	0.407	761.5	$0.375	83	128,000
72	gpt-oss-120b · OpenAI	30.4	24.1	0.235	0.510	163.3	$0.065	468	131,072
73	Qwen3.5-9B · Alibaba / Qwen	28.7	21.8	0.242	0.653	84.6	$0.112	255	262,144
74	Command A · Cohere	27.8	22.8	0.250	0.487	192.5	$4.375	6	256,000
75	Nemotron 3.5 Lightning · Nvidia	26.8	23.6	—	0.553	287.6	$0.138	195	1,000,000
76	Mistral Small 4 · Mistral	26.6	19.7	0.174	0.473	144.2	$0.262	101	262,144
77	Ling 3.0 Tiny · InclusionAI	26.5	24.5	—	0.587	228.2	—	—	—
78	Mistral Small 3.1 · Mistral	26.3	14.9	0.076	0.220	0.0	$0.402	65	128,000
79	Trinity Large Thinking · Arcee Ai	25.8	18.7	0.227	0.383	212.3	$0.378	68	262,144
80	DeepSeek R1 · DeepSeek	24.6	18.6	0.061	0.560	0.0	$1.150	21	64,000
81	GPT-4o · OpenAI	24.2	8.4	—	—	0.0	$4.375	6	128,000
82	DeepSeek V3 · DeepSeek	23.0	14.2	0.068	0.317	0.0	$0.450	51	163,840
83	Qwen3 235B A22B 2507 · Alibaba / Qwen	22.1	19.9	0.136	0.707	0.0	$0.796	28	131,072
84	GPT-4 Turbo · OpenAI	21.5	7.7	—	—	0.0	$15.000	1	128,000
85	DeepSeek V3 0324 · DeepSeek	21.2	15.2	0.152	0.413	0.0	$0.483	44	163,840
86	gpt-oss-20b · OpenAI	20.7	15.2	0.106	0.333	125.6	$0.055	376	131,072
87	Mistral Medium 3.1 · Mistral	20.5	14.7	0.106	0.213	0.0	$0.800	26	131,072
88	GPT-4.1 Mini · OpenAI	20.2	14.8	0.076	0.453	0.0	$0.700	29	1,047,576
89	Mistral Large 3 · Mistral	20.1	15.9	0.159	0.347	56.5	$3.000	7	128,000
90	Llama 4 Maverick · Meta	16.3	14.5	0.068	0.500	112.9	$0.350	47	1,048,576
91	o3 Mini · OpenAI	16.3	15.7	0.061	0.420	0.0	$1.925	8	200,000
92	Solar Pro 3 · Upstage	16.2	14.5	0.076	0.310	152.8	$0.262	62	131,072
93	GPT-5 Mini · OpenAI	15.6	25.8	0.333	0.710	0.0	$0.688	23	400,000
94	Qwen3 32B · Alibaba / Qwen	15.3	11.4	0.030	0.000	0.0	$0.130	118	131,072
95	Nemotron 3 Nano 30B · Nvidia	14.4	14.5	0.136	0.373	253.6	$0.087	165	262,144
96	Qwen3 14B · Alibaba / Qwen	13.8	10.4	0.038	0.000	0.0	$0.150	92	131,072
97	Nemotron 3 Nano Omni 30B · Nvidia	13.8	15.0	0.083	0.407	328.6	$0.000	—	256,000
98	GPT-4 · OpenAI	13.1	6.8	—	—	0.0	$37.500	0	8,191
99	Mistral Small 3.2 · Mistral	12.5	10.7	0.068	0.187	0.0	$0.133	94	256,000
100	Qwen3 30B A3B 2507 · Alibaba / Qwen	12.1	14.6	0.053	0.597	0.0	$0.215	56	131,072
101	GPT-4o-mini · OpenAI	11.4	6.7	—	—	0.0	$0.262	43	128,000
102	GPT-4.1 Nano · OpenAI	11.1	9.6	0.038	0.193	0.0	$0.175	63	1,047,576
103	GPT-3.5 Turbo · OpenAI	10.7	3.2	—	—	0.0	$0.750	14	16,385
104	Gemma 3 27B · Google	10.1	7.4	0.038	0.063	0.0	$0.172	59	262,144
105	Granite 4.1 8B · IBM	9.5	6.4	0.000	0.113	128.6	$0.062	152	131,072
106	Qwen3 8B · Alibaba / Qwen	9.0	8.3	0.023	0.000	0.0	$0.202	45	131,072
107	Llama 4 Scout · Meta	8.2	10.3	0.015	0.303	131.1	$0.150	55	1,310,720
108	Gemma 3 12B · Google	5.8	5.5	0.008	0.080	0.0	$0.075	77	131,072
109	Gemma 3n E4B · Google	3.2	1.0	0.023	0.000	0.0	$0.075	43	32,768
110	Gemma 3 4B · Google	2.7	1.0	0.008	0.067	0.0	$0.062	43	131,072
—	Nemotron 3 Super 120B · Nvidia	—	—	—	—	—	$0.164	—	1,000,000

Quality and speed are third-party measurements from Artificial Analysis — the only source publishing measured speed. A dash means the model has not been independently measured, not that it scored zero. Intelligence is AA’s general index, shown next to the coding one because they diverge: a model can code well and reason poorly, or the reverse, and only the coding column decides the ranking here. Long-ctx is AA’s long-context reasoning score — the column that matters on a large codebase, since Context only says how many tokens fit, not whether the model still reasons at that length.

Published quality

Mind the provenance

The flattering coding scores are vendor-reported. Where an independent lab measured the same benchmark the numbers agree closely — so the gap is about which benchmarks get published, not fabricated values.

Model	Benchmark	Score	Provenance	Source
MiniMax-M3	SWE-bench Verified	80.5%	vendor	minimax.io
MiniMax-M3	SWE-bench Pro	59.0%	vendor	minimax.io
MiniMax-M3	Terminal-Bench 2.1	66.0%	vendor	minimax.io
MiniMax-M3	Terminal-Bench 2.1	65.2%	independent	Artificial Analysis
MiniMax-M3	TerminalBench-Hard	42.4%	independent	Artificial Analysis
MiniMax-M3	Aider polyglot	not listed	independent	aider.chat
Xiaomi MiMo-V2.5-Pro	SWE-bench Verified	78.9%	vendor	mimo.xiaomi.com
Mistral Devstral Small 2	SWE-bench Verified	68.0%	vendor	mistral.ai

Models the coding index can’t rank

Measured here, not by AA

The leaderboard ranks on Artificial Analysis’ coding index — a proprietary composite AA publishes only for models it chooses to test. Some models here have no such score and cannot get one from us: the index is not a benchmark we can run. So they were measured on tests we can run, and placed next to models that appear on both scales.

BigCodeBench-Hard saturation - the SAME 148 problems, every model comparable

Model	Score	AA coding index
Granite 4.1 8B	17%	9.5
Mistral Small 3.2	25%	12.5
Ministral 3 14B	14%	14.4
Nemotron Nano 9B V2	22%	not measured
gpt-oss-20b	26%	20.7
North Mini Code	35%	36.5
Nemotron 3 Super 120B	38%	37.7
Gemma 4 26B A4B	36%	39.3
Nemotron 3 Ultra 550B	44%	49.3
DeepSeek V4 Flash	33%	56.2
GLM-5.2	39%	68.8

BigCodeBench-Hard (run here)

Model	Score	AA coding index
MiniMax	40.0%	58.6
Agnes AI	35.0%	not measured

SWE-bench Verified (run here, 25 instances, 80-step budget)

Model	Score	AA coding index
Tencent	68% (17/25)	58.8
Agnes AI	52% (13/25)	not measured

The saturation table is the honest shape of our own BigCodeBench-Hard, run on ONE fixed set of 148 problems so every model is directly comparable (an earlier pass compared different problem subsets and was invalid). It climbs with the coding index up to about index 49 (~44%), then FLATTENS: DeepSeek V4 Flash (index 56) and GLM-5.2 (index 69) land at 33-39%, no better than a 550B Nemotron at 49. It does NOT rise to 100% - the ceiling is ~40-44%, because 36% of these problems (44 of 121 that all models attempt) are passed by NOBODY: 38 are genuine wrong answers, only 6 are offline-sandbox errors. So the ceiling is real (hard problems + strict tests that reject valid alternative solutions), not a grading artefact. The lesson for this site: our BCB-Hard discriminates coders up to ~index 45; above that it saturates and the AA coding index is the better ranker - which is exactly how the local picks are ordered. Grades are offline (--network none), which caps everyone equally, so read absolute scores as a floor. Below, older single-benchmark brackets (SWE-bench Verified, 25 instances) - read as a bracket, not a coordinate: across those gold-gated instances Tencent Hy3 (index 58.8) resolved 68% and Agnes 52%, but most of that gap is our test - Agnes failed to converge in the 80-step budget on 9 of 25 vs 1 for Hy3, and 2 of 4 re-run at 200 steps then resolved. These stay off the leaderboard and never become an index.

Providers & quantization

The column nobody else shows

The same model is served by many providers at nearly the same price but at different numeric precision. Cheapest is often the most aggressively quantized — and sometimes higher precision costs the same or less.

bf16fp8fp4unknown

GPT-5.6 Sol

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$2.50	$15.00	1,050,000	99.6%
Azure	unknown	$5.00	$30.00	1,050,000	99.9%
OpenAI	unknown	$5.00	$30.00	1,050,000	99.6%
Amazon Bedrock	unknown	$5.50	$33.00	1,050,000	—
Azure	unknown	$5.50	$33.00	1,050,000	100.0%
Azure	unknown	$5.50	$33.00	1,050,000	—
OpenAI	unknown	$10.00	$60.00	1,050,000	99.6%

Claude Opus 5

10 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Amazon Bedrock	unknown	$5.00	$25.00	1,000,000	3.7%
Amazon Bedrock	unknown	$5.00	$25.00	1,000,000	99.4%
Anthropic	unknown	$5.00	$25.00	1,000,000	98.5%
Azure	unknown	$5.00	$25.00	1,000,000	100.0%
Azure	unknown	$5.00	$25.00	1,000,000	—
Claude Platform on AWS	unknown	$5.00	$25.00	1,000,000	—
Google	unknown	$5.00	$25.00	1,000,000	100.0%
Amazon Bedrock	unknown	$5.50	$27.50	1,000,000	100.0%
Google	unknown	$5.50	$27.50	1,000,000	100.0%
Google	unknown	$5.50	$27.50	1,000,000	—

GPT-5.6 Terra

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.50	$3.00	1,050,000	99.9%
OpenAI	unknown	$1.00	$6.00	1,050,000	99.9%
Azure	unknown	$2.00	$12.00	1,050,000	99.9%
OpenAI	unknown	$2.00	$12.00	1,050,000	99.9%
Amazon Bedrock	unknown	$2.20	$13.20	1,050,000	—
Azure	unknown	$2.20	$13.20	1,050,000	100.0%
Azure	unknown	$2.20	$13.20	1,050,000	—

Claude Fable 5

6 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Anthropic direct	first-party	$10.00	$50.00	—	—
Amazon Bedrock	unknown	$10.00	$50.00	1,000,000	—
Amazon Bedrock	unknown	$10.00	$50.00	1,000,000	99.6%
Anthropic	unknown	$10.00	$50.00	1,000,000	100.0%
Azure	unknown	$10.00	$50.00	1,000,000	—
Google	unknown	$10.00	$50.00	1,000,000	100.0%
Google	unknown	$11.00	$55.00	1,000,000	—

Buying direct from Anthropic costs $10.00 in and $50.00 out per million, cached input $1.000.

Kimi K3

14 providers · bf16, fp4, fp8, mxfp4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Moonshot direct	first-party	$3.00	$15.00	—	—
Morph	fp4	$2.80	$14.00	1,048,576	99.9%
DeepInfra	bf16	$2.85	$14.25	1,048,576	—
DigitalOcean	unknown	$2.85	$14.25	1,048,576	99.8%
BaseTen	fp8	$3.00	$15.00	1,048,576	99.0%
Chutes	mxfp4	$3.00	$15.00	1,048,576	98.1%
Fireworks	unknown	$3.00	$15.00	1,048,576	99.6%
Modal	mxfp4	$3.00	$15.00	1,048,576	98.9%
Moonshot	mxfp4	$3.00	$15.00	1,048,576	100.0%
Phala	unknown	$3.00	$15.00	1,048,576	99.4%
Sail Research	fp4	$3.00	$15.00	974,842	99.6%
Together	unknown	$3.00	$15.00	1,000,000	99.1%
Wafer	unknown	$3.00	$15.00	1,048,576	98.4%
Fireworks	unknown	$4.50	$22.50	1,048,576	99.6%
Morph	fp4	$6.00	$22.50	1,048,576	—

Buying direct from Moonshot costs $3.00 in and $15.00 out per million, cached input $0.300. That is 7% dearer than the cheapest routed endpoint on a 3:1 blend.

GPT-5.5

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$2.50	$15.00	1,050,000	99.9%
Azure	unknown	$5.00	$30.00	1,050,000	99.9%
OpenAI	unknown	$5.00	$30.00	1,050,000	99.9%
Amazon Bedrock	unknown	$5.50	$33.00	1,050,000	—
Azure	unknown	$5.50	$33.00	1,050,000	—
Azure	unknown	$5.50	$33.00	1,050,000	100.0%
OpenAI	unknown	$12.50	$75.00	1,050,000	99.9%

Claude Opus 4.8

10 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Anthropic direct	first-party	$5.00	$25.00	—	—
Amazon Bedrock	unknown	$5.00	$25.00	1,000,000	100.0%
Anthropic	unknown	$5.00	$25.00	1,000,000	99.7%
Anthropic	unknown	$5.00	$25.00	1,000,000	99.9%
Azure	unknown	$5.00	$25.00	1,000,000	—
Azure	unknown	$5.00	$25.00	1,000,000	—
Google	unknown	$5.00	$25.00	1,000,000	100.0%
Amazon Bedrock	unknown	$5.50	$27.50	1,000,000	—
Amazon Bedrock	unknown	$5.50	$27.50	1,000,000	—
Google	unknown	$5.50	$27.50	1,000,000	—
Google	unknown	$5.50	$27.50	1,000,000	—

Buying direct from Anthropic costs $5.00 in and $25.00 out per million, cached input $0.500.

Claude Opus 4.7

8 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Anthropic direct	first-party	$5.00	$25.00	—	—
Amazon Bedrock	unknown	$5.00	$25.00	1,000,000	100.0%
Anthropic	unknown	$5.00	$25.00	1,000,000	97.2%
Anthropic	unknown	$5.00	$25.00	1,000,000	95.6%
Azure	unknown	$5.00	$25.00	1,000,000	—
Google	unknown	$5.00	$25.00	1,000,000	100.0%
Amazon Bedrock	unknown	$5.50	$27.50	1,000,000	—
Google	unknown	$5.50	$27.50	1,000,000	—
Google	unknown	$5.50	$27.50	1,000,000	—

Buying direct from Anthropic costs $5.00 in and $25.00 out per million, cached input $0.500.

Grok 4.5

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
xAI direct	first-party	$2.00	$6.00	—	—
xAI	unknown	$2.00	$6.00	500,000	100.0%
xAI	unknown	$2.00	$6.00	500,000	99.8%
xAI	unknown	$4.00	$12.00	500,000	100.0%
xAI	unknown	$4.00	$12.00	500,000	99.8%

Buying direct from xAI costs $2.00 in and $6.00 out per million, cached input $0.300.

Grok 4.6

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
xAI	unknown	$2.00	$6.00	500,000	99.7%
xAI	unknown	$2.00	$6.00	500,000	99.8%
xAI	unknown	$4.00	$12.00	500,000	99.7%
xAI	unknown	$4.00	$12.00	500,000	99.8%

Muse Glimmer

4 providers · bf16, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	bf16	$0.30	$1.20	131,072	100.0%
Fireworks	unknown	$0.35	$1.50	131,072	100.0%
Phala	unknown	$0.35	$1.50	131,072	97.2%
Together	unknown	$0.35	$1.50	131,072	99.7%

Not listed above: Meta directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Solar Pro 4

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Upstage	unknown	$0.03	$0.12	524,288	100.0%

Nemotron 3.5 Lightning

3 providers · bf16, fp4

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	bf16	$0.08	$0.20	262,144	99.6%
CoreWeave	bf16	$0.10	$0.25	262,144	100.0%
Venice	fp4	$0.10	$0.25	1,000,000	100.0%

Not listed above: Nvidia directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Claude Sonnet 5

9 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Anthropic direct	first-party	$2.00	$10.00	—	—
Amazon Bedrock	unknown	$2.00	$10.00	1,000,000	100.0%
Amazon Bedrock	unknown	$2.00	$10.00	1,000,000	100.0%
Anthropic	unknown	$2.00	$10.00	1,000,000	100.0%
Azure	unknown	$2.00	$10.00	1,000,000	100.0%
Azure	unknown	$2.00	$10.00	1,000,000	—
Google	unknown	$2.00	$10.00	1,000,000	100.0%
Amazon Bedrock	unknown	$2.20	$11.00	1,000,000	—
Google	unknown	$2.20	$11.00	1,000,000	100.0%
Google	unknown	$2.20	$11.00	1,000,000	—

Buying direct from Anthropic costs $2.00 in and $10.00 out per million, cached input $0.200.

GPT-5.6 Luna

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.05	$0.30	1,050,000	99.9%
OpenAI	unknown	$0.10	$0.60	1,050,000	99.9%
Azure	unknown	$0.20	$1.20	1,050,000	100.0%
OpenAI	unknown	$0.20	$1.20	1,050,000	99.9%
Amazon Bedrock	unknown	$0.22	$1.32	1,050,000	100.0%
Azure	unknown	$0.22	$1.32	1,050,000	100.0%
Azure	unknown	$0.22	$1.32	1,050,000	100.0%

Muse Spark 1.1

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Meta	unknown	$1.25	$4.25	1,048,576	100.0%

GPT-5.4

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$1.25	$7.50	1,050,000	76.0%
Azure	unknown	$2.50	$15.00	1,050,000	100.0%
OpenAI	unknown	$2.50	$15.00	1,050,000	76.0%
Amazon Bedrock	unknown	$2.75	$16.50	1,050,000	—
Azure	unknown	$2.75	$16.50	1,050,000	100.0%
Azure	unknown	$2.75	$16.50	1,050,000	100.0%
OpenAI	unknown	$5.00	$30.00	1,050,000	76.0%

Gemini 3.5 Flash

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$0.75	$4.50	1,048,576	99.9%
Google AI Studio	unknown	$0.75	$4.50	1,048,576	99.7%
Google	unknown	$1.50	$9.00	1,048,576	99.9%
Google AI Studio	unknown	$1.50	$9.00	1,048,576	99.7%
Google	unknown	$1.65	$9.90	1,048,576	—
Google	unknown	$2.70	$16.20	1,048,576	99.9%
Google AI Studio	unknown	$2.70	$16.20	1,048,576	99.7%

Gemini 3.6 Flash

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$0.38	$1.88	1,048,576	99.8%
Google AI Studio	unknown	$0.38	$1.88	1,048,576	99.7%
Google	unknown	$0.75	$3.75	1,048,576	99.8%
Google AI Studio	unknown	$0.75	$3.75	1,048,576	99.7%
Google	unknown	$0.83	$4.12	1,048,576	—
Google	unknown	$1.35	$6.75	1,048,576	99.8%
Google AI Studio	unknown	$1.35	$6.75	1,048,576	99.7%

GLM-5.2

32 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Z.ai / Zhipu direct	first-party	$1.40	$4.40	—	—
Baidu	fp8	$0.49	$1.54	1,048,576	100.0%
DigitalOcean	unknown	$0.63	$1.98	262,144	100.0%
Novita	fp8	$0.74	$2.33	1,048,576	99.3%
GMICloud	fp8	$0.74	$2.33	1,048,576	98.9%
StreamLake	fp8	$0.75	$2.37	1,024,000	99.7%
DeepInfra	fp4	$0.75	$2.40	1,048,576	99.6%
Sail Research	fp8	$0.50	$3.15	1,048,576	99.6%
CoreWeave	fp4	$0.76	$2.42	262,144	99.9%
AkashML	fp8	$0.77	$2.42	96,890	96.7%
Decart	fp4	$0.77	$2.56	1,048,576	99.9%
Inceptron	fp4	$0.75	$2.90	1,048,576	98.8%
Alibaba / Qwen	fp8	$0.97	$3.04	1,048,576	100.0%
Phala	fp8	$1.13	$3.00	1,048,576	99.7%
SiliconFlow	fp8	$1.19	$3.74	1,048,576	100.0%
Morph	fp4	$1.10	$4.10	1,048,576	96.1%
Ambient	fp8	$1.05	$4.40	202,752	98.5%
AtlasCloud	fp8	$1.26	$3.96	1,048,576	100.0%
Wafer	fp4	$1.26	$3.96	1,048,576	98.5%
BaseTen	fp8	$1.40	$4.40	1,048,576	100.0%
Cloudflare	unknown	$1.40	$4.40	262,144	—
Crusoe	fp8	$1.40	$4.40	1,048,576	97.6%
Fireworks	unknown	$1.40	$4.40	1,048,576	98.7%
Friendli	unknown	$1.40	$4.40	1,048,576	98.6%
Parasail	fp4	$1.40	$4.40	262,144	99.9%
Together	unknown	$1.40	$4.40	512,000	99.1%
Venice	fp8	$1.40	$4.40	1,000,000	98.9%
Z.ai / Zhipu	fp8	$1.40	$4.40	1,048,576	99.9%
BaseTen	fp8	$2.10	$6.60	1,048,576	100.0%
Cloudflare	unknown	$2.10	$6.60	262,144	100.0%
Fireworks	unknown	$2.10	$6.60	1,048,576	98.5%
Wafer	fp4	$2.10	$6.60	1,048,576	99.6%
Alibaba / Qwen	fp8	$2.31	$7.26	1,048,576	100.0%

Buying direct from Z.ai / Zhipu costs $1.40 in and $4.40 out per million, cached input $0.260. That is 186% dearer than the cheapest routed endpoint on a 3:1 blend.

Gemini 3.1 Pro Preview

6 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$1.00	$6.00	1,048,576	97.2%
Google AI Studio	unknown	$1.00	$6.00	1,048,576	99.8%
Google	unknown	$2.00	$12.00	1,048,576	97.2%
Google AI Studio	unknown	$2.00	$12.00	1,048,576	99.8%
Google	unknown	$3.60	$21.60	1,048,576	97.2%
Google AI Studio	unknown	$3.60	$21.60	1,048,576	99.8%

Qwen3.7 Max

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$2.50	$7.50	—	—
Alibaba / Qwen	unknown	$1.48	$4.42	1,000,000	100.0%

Buying direct from Alibaba / Qwen costs $2.50 in and $7.50 out per million, cached input $0.500. That is 69% dearer than the cheapest routed endpoint on a 3:1 blend.

Claude Sonnet 4.6

9 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Anthropic direct	first-party	$3.00	$15.00	—	—
Amazon Bedrock	unknown	$3.00	$15.00	1,000,000	99.9%
Anthropic	unknown	$3.00	$15.00	1,000,000	100.0%
Anthropic	unknown	$3.00	$15.00	1,000,000	—
Azure	unknown	$3.00	$15.00	1,000,000	—
Google	unknown	$3.00	$15.00	1,000,000	99.9%
Amazon Bedrock	unknown	$3.30	$16.50	1,000,000	—
Amazon Bedrock	unknown	$3.30	$16.50	1,000,000	—
Google	unknown	$3.30	$16.50	1,000,000	—
Google	unknown	$3.30	$16.50	1,000,000	—

Buying direct from Anthropic costs $3.00 in and $15.00 out per million, cached input $0.300.

Kimi K2.6

21 providers · bf16, fp4, fp8, int4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Moonshot direct	first-party	$0.95	$4.00	—	—
Baidu	fp4	$0.56	$2.36	262,144	100.0%
Decart	fp4	$0.57	$2.39	262,144	99.6%
StreamLake	fp8	$0.60	$2.52	256,000	98.5%
Chutes	int4	$0.58	$3.40	262,144	99.6%
Inceptron	int4	$0.60	$3.41	262,144	100.0%
CoreWeave	fp4	$0.65	$3.41	262,144	99.9%
DigitalOcean	unknown	$0.76	$3.20	262,144	99.6%
Crusoe	bf16	$0.70	$3.50	262,144	88.2%
SiliconFlow	fp8	$0.77	$3.40	262,144	99.7%
DeepInfra	fp4	$0.75	$3.50	262,144	100.0%
Parasail	int4	$0.75	$3.50	262,144	99.6%
Venice	int4	$0.75	$3.50	256,000	81.4%
Novita	unknown	$0.80	$3.40	262,144	99.9%
AtlasCloud	int4	$0.95	$4.00	262,144	91.1%
BaseTen	fp4	$0.95	$4.00	262,000	—
Cloudflare	unknown	$0.95	$4.00	262,144	100.0%
Fireworks	unknown	$0.95	$4.00	262,144	—
Moonshot	int4	$0.95	$4.00	262,144	100.0%
Sail Research	int4	$1.00	$4.00	262,144	100.0%
Phala	unknown	$1.09	$4.60	262,144	96.3%
Together	unknown	$1.20	$4.50	262,144	95.0%

Buying direct from Moonshot costs $0.95 in and $4.00 out per million, cached input $0.160. That is 69% dearer than the cheapest routed endpoint on a 3:1 blend.

Kimi K2.7 Code

15 providers · fp4, fp8, int4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Moonshot direct	first-party	$0.95	$4.00	—	—
Inceptron	int4	$0.67	$3.40	262,144	99.9%
DeepInfra	fp4	$0.68	$3.40	262,144	100.0%
Ambient	unknown	$0.69	$3.49	262,144	99.7%
CoreWeave	int4	$0.71	$3.50	262,144	100.0%
Venice	int4	$0.75	$3.50	256,000	100.0%
Parasail	int4	$0.76	$3.50	262,144	99.8%
ModelRun	fp4	$0.85	$3.75	262,144	99.7%
SiliconFlow	fp8	$0.86	$3.80	262,144	100.0%
Novita	int4	$0.91	$3.84	262,144	100.0%
Alibaba / Qwen	fp8	$0.95	$4.00	262,144	100.0%
AtlasCloud	int4	$0.95	$4.00	262,144	—
Cloudflare	unknown	$0.95	$4.00	262,144	—
Moonshot	int4	$0.95	$4.00	262,144	100.0%
Together	unknown	$0.95	$4.00	262,144	98.7%
Moonshot	int4	$1.90	$8.00	262,144	100.0%

Buying direct from Moonshot costs $0.95 in and $4.00 out per million, cached input $0.190. That is 27% dearer than the cheapest routed endpoint on a 3:1 blend.

Xiaomi MiMo-V2.5-Pro

7 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Xiaomi direct	first-party	$0.43	$0.87	—	—
AtlasCloud	fp8	$0.43	$0.87	1,024,000	100.0%
GMICloud	bf16	$0.43	$0.87	1,050,000	98.2%
Xiaomi	fp8	$0.43	$0.87	1,048,576	100.0%
Novita	unknown	$0.48	$0.96	1,048,576	100.0%
StreamLake	unknown	$0.52	$1.04	1,000,000	—
DigitalOcean	unknown	$0.40	$1.50	262,144	100.0%
DeepInfra	fp8	$1.00	$3.00	1,048,576	100.0%

Buying direct from Xiaomi costs $0.43 in and $0.87 out per million, cached input $0.004.

KAT-Coder-Pro V2

2 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
AtlasCloud	fp8	$0.30	$1.20	262,144	100.0%
StreamLake	unknown	$0.30	$1.20	256,000	—

Not listed above: Kwaipilot directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

DeepSeek V4 Pro

7 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepSeek direct	first-party	$0.43	$0.87	—	—
DeepSeek	unknown	$0.43	$0.87	1,048,576	100.0%
GMICloud	fp8	$1.22	$2.44	1,048,575	99.9%
BaseTen	fp4	$1.32	$3.96	1,048,576	100.0%
Cloudflare	unknown	$1.32	$3.96	1,048,576	99.8%
Fireworks	unknown	$1.32	$3.96	1,048,576	100.0%
Novita	fp8	$1.32	$3.96	1,048,576	99.7%
SiliconFlow	fp8	$1.32	$3.96	1,048,576	100.0%

Buying direct from DeepSeek costs $0.43 in and $0.87 out per million, cached input $0.004.

Nex-N2-Pro

2 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Nex AGI	fp8	$0.25	$1.00	262,144	—
SiliconFlow	unknown	$0.50	$2.50	262,144	—

Tencent Hy3

6 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Tencent direct	first-party	$0.15	$0.59	—	—
Baidu	fp8	$0.13	$0.51	262,144	99.9%
GMICloud	bf16	$0.13	$0.53	262,144	99.2%
Tencent	fp8	$0.13	$0.53	262,144	99.9%
DeepInfra	fp8	$0.14	$0.58	262,144	96.5%
Novita	unknown	$0.14	$0.58	262,144	100.0%
AtlasCloud	fp8	$0.20	$0.80	262,144	99.8%

Buying direct from Tencent costs $0.15 in and $0.59 out per million, cached input $0.037. That is 16% dearer than the cheapest routed endpoint on a 3:1 blend. Priced in RMB: 1 / 4 / 0.25 per million (input / output / cached), converted at about 6.8 CNY to the dollar.

MiniMax-M3

12 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
MiniMax direct	first-party	$0.30	$1.20	—	—
CoreWeave	fp4	$0.23	$0.96	262,144	99.9%
GMICloud	fp8	$0.24	$0.96	1,048,576	97.8%
DeepInfra	fp8	$0.28	$1.10	524,288	98.8%
AtlasCloud	fp8	$0.30	$1.20	524,300	99.7%
MiniMax	fp8	$0.30	$1.20	524,288	97.6%
Morph	fp4	$0.30	$1.20	256,000	96.7%
Novita	fp8	$0.30	$1.20	1,000,000	99.7%
Parasail	fp8	$0.30	$1.20	1,048,576	99.0%
StreamLake	fp8	$0.30	$1.20	1,000,000	99.5%
Together	unknown	$0.30	$1.20	524,288	99.5%
Venice	fp8	$0.30	$1.20	524,288	99.8%
ModelRun	fp4	$0.75	$3.00	1,048,576	99.9%

Buying direct from MiniMax costs $0.30 in and $1.20 out per million, cached input $0.060. That is 27% dearer than the cheapest routed endpoint on a 3:1 blend.

Xiaomi MiMo-V2.5

5 providers · bf16, fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Xiaomi direct	first-party	$0.14	$0.28	—	—
GMICloud	fp8	$0.14	$0.28	1,050,000	99.5%
Parasail	fp8	$0.14	$0.28	1,048,576	99.4%
Xiaomi	fp8	$0.14	$0.28	1,048,576	99.8%
Novita	fp8	$0.17	$0.34	1,048,576	99.1%
DeepInfra	bf16	$0.40	$2.00	262,144	97.2%

Buying direct from Xiaomi costs $0.14 in and $0.28 out per million, cached input $0.003.

DeepSeek V4 Flash

28 providers · bf16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepSeek direct	first-party	$0.14	$0.28	—	—
Baidu	fp8	$0.08	$0.16	1,048,576	90.4%
Decart	fp4	$0.08	$0.16	262,144	99.9%
GMICloud	fp8	$0.08	$0.17	1,048,575	99.7%
DeepInfra	fp4	$0.08	$0.18	1,048,576	99.1%
OpenInference	fp4	$0.08	$0.18	262,144	96.2%
Sail Research	fp4	$0.09	$0.18	262,144	99.4%
StreamLake	fp8	$0.09	$0.18	1,024,000	99.8%
DigitalOcean	unknown	$0.08	$0.25	1,048,576	99.1%
BaseTen	fp8	$0.13	$0.26	1,048,576	99.7%
CoreWeave	fp8	$0.13	$0.28	262,144	100.0%
Inceptron	fp4	$0.13	$0.28	1,048,576	99.8%
Morph	bf16	$0.14	$0.28	1,048,576	98.5%
AkashML	fp8	$0.14	$0.28	131,072	99.5%
Ambient	fp4	$0.14	$0.28	1,048,576	97.9%
AtlasCloud	fp4	$0.14	$0.28	1,048,576	99.5%
Cloudflare	unknown	$0.14	$0.28	1,048,576	100.0%
DeepSeek	fp8	$0.14	$0.28	1,048,576	100.0%
Fireworks	unknown	$0.14	$0.28	1,048,576	96.4%
Novita	fp8	$0.14	$0.28	1,048,576	99.7%
Parasail	fp8	$0.14	$0.28	1,048,576	94.9%
Relace	fp4	$0.14	$0.28	1,048,576	99.5%
SiliconFlow	fp8	$0.14	$0.28	1,048,576	99.8%
Together	unknown	$0.14	$0.28	1,048,576	97.1%
Io Net	fp8	$0.15	$0.32	262,100	99.8%
Venice	unknown	$0.17	$0.35	1,000,000	97.0%
Phala	unknown	$0.20	$0.40	1,048,576	99.3%
Mancer 2	fp8	$0.17	$0.50	1,048,576	97.6%
Wafer	unknown	$0.28	$0.56	1,048,576	99.3%

Buying direct from DeepSeek costs $0.14 in and $0.28 out per million, cached input $0.003. That is 75% dearer than the cheapest routed endpoint on a 3:1 blend.

GPT-5.4 Nano

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.10	$0.62	400,000	85.5%
Azure	unknown	$0.20	$1.25	400,000	100.0%
OpenAI	unknown	$0.20	$1.25	400,000	85.5%
Azure	unknown	$0.22	$1.38	400,000	100.0%

GPT-5.4 Mini

5 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.38	$2.25	400,000	100.0%
Azure	unknown	$0.75	$4.50	400,000	100.0%
OpenAI	unknown	$0.75	$4.50	400,000	100.0%
Azure	unknown	$0.83	$4.95	400,000	100.0%
OpenAI	unknown	$1.50	$9.00	400,000	100.0%

Qwen3.7 Plus

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.50	$3.00	—	—
Alibaba / Qwen	unknown	$0.32	$1.28	1,000,000	100.0%

Buying direct from Alibaba / Qwen costs $0.50 in and $3.00 out per million, cached input $0.050. That is 101% dearer than the cheapest routed endpoint on a 3:1 blend.

GLM-5.1

18 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Z.ai / Zhipu direct	first-party	$1.40	$4.40	—	—
StreamLake	fp8	$0.97	$3.04	200,000	99.8%
Chutes	fp8	$0.98	$3.08	202,752	74.6%
Wafer	fp4	$1.00	$3.20	202,752	95.5%
DeepInfra	fp4	$1.05	$3.50	202,752	100.0%
DigitalOcean	unknown	$0.97	$4.30	163,840	92.6%
SiliconFlow	fp8	$1.19	$3.74	204,800	98.8%
AtlasCloud	fp8	$1.26	$3.96	202,752	100.0%
Phala	unknown	$1.21	$4.20	202,752	—
Crusoe	fp8	$1.20	$4.40	202,752	100.0%
Alibaba / Qwen	fp8	$1.33	$4.18	202,745	99.5%
Novita	fp8	$1.38	$4.40	204,800	100.0%
Baidu	fp8	$1.40	$4.40	202,752	100.0%
Friendli	unknown	$1.40	$4.40	202,752	100.0%
GMICloud	fp8	$1.40	$4.40	202,752	99.3%
Nebius	fp8	$1.40	$4.40	202,752	99.3%
Parasail	fp8	$1.40	$4.40	202,752	100.0%
Z.ai / Zhipu	fp8	$1.40	$4.40	202,752	99.7%
Venice	fp8	$1.54	$4.84	200,000	99.4%

Buying direct from Z.ai / Zhipu costs $1.40 in and $4.40 out per million, cached input $0.260. That is 45% dearer than the cheapest routed endpoint on a 3:1 blend.

Qwen3.6 Plus

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.50	$3.00	—	—
Alibaba / Qwen	unknown	$0.33	$1.95	1,000,000	100.0%

Buying direct from Alibaba / Qwen costs $0.50 in and $3.00 out per million, cached input $0.050. That is 54% dearer than the cheapest routed endpoint on a 3:1 blend.

Qwen3.6 27B

9 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.60	$3.60	—	—
Chutes	fp8	$0.30	$2.00	262,144	52.7%
Morph	fp4	$0.29	$2.40	131,072	100.0%
Phala	unknown	$0.32	$2.70	262,144	95.5%
Alibaba / Qwen	unknown	$0.45	$2.70	262,144	100.0%
Io Net	fp8	$0.39	$2.89	32,768	100.0%
SiliconFlow	fp8	$0.30	$3.20	262,144	99.9%
DeepInfra	fp8	$0.32	$3.20	262,144	99.9%
Venice	fp8	$0.33	$3.25	256,000	100.0%
CoreWeave	fp8	$0.60	$3.60	262,144	99.8%

Buying direct from Alibaba / Qwen costs $0.60 in and $3.60 out per million. That is 86% dearer than the cheapest routed endpoint on a 3:1 blend.

MiniMax M2.7

11 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
MiniMax direct	first-party	$0.30	$1.20	—	—
Mara	unknown	$0.24	$0.96	196,608	100.0%
DeepInfra	fp8	$0.25	$1.00	196,608	100.0%
GMICloud	fp8	$0.27	$1.08	196,608	100.0%
Novita	fp8	$0.27	$1.08	204,800	98.2%
AtlasCloud	fp8	$0.30	$1.20	196,608	—
Fireworks	unknown	$0.30	$1.20	196,608	—
MiniMax	fp8	$0.30	$1.20	204,800	97.6%
DeepInfra	fp8	$0.38	$1.70	196,608	99.5%
Groq	unknown	$0.60	$1.80	196,608	100.0%
MiniMax	fp8	$0.60	$2.40	204,800	—
SambaNova	unknown	$0.60	$2.40	196,608	—

Buying direct from MiniMax costs $0.30 in and $1.20 out per million, cached input $0.060. That is 25% dearer than the cheapest routed endpoint on a 3:1 blend.

Inkling

3 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp8	$0.95	$4.05	524,288	99.8%
BaseTen	fp8	$1.00	$4.05	1,048,576	99.8%
Together	unknown	$1.00	$4.05	524,288	100.0%

Not listed above: Thinking Machines directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

GPT-5.1

5 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.62	$5.00	400,000	99.6%
Azure	unknown	$1.25	$10.00	400,000	100.0%
OpenAI	unknown	$1.25	$10.00	400,000	99.6%
Azure	unknown	$1.38	$11.00	400,000	—
OpenAI	unknown	$2.50	$20.00	400,000	99.6%

Gemini 3.5 Flash-Lite

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$0.15	$1.25	1,048,576	99.9%
Google AI Studio	unknown	$0.15	$1.25	1,048,576	99.6%
Google	unknown	$0.30	$2.50	1,048,576	99.9%
Google AI Studio	unknown	$0.30	$2.50	1,048,576	99.6%
Google	unknown	$0.33	$2.75	1,048,576	—
Google	unknown	$0.54	$4.50	1,048,576	99.9%
Google AI Studio	unknown	$0.54	$4.50	1,048,576	99.6%

Qwen3.5 397B A17B

11 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.60	$3.60	—	—
DigitalOcean	unknown	$0.30	$1.93	131,072	100.0%
Alibaba / Qwen	unknown	$0.39	$2.34	262,144	100.0%
Chutes	fp8	$0.45	$3.00	262,144	99.7%
DeepInfra	fp8	$0.45	$3.00	262,144	99.8%
Parasail	fp8	$0.50	$3.60	262,144	99.9%
AtlasCloud	fp8	$0.55	$3.50	262,144	99.3%
Phala	unknown	$0.55	$3.50	262,144	100.0%
GMICloud	fp8	$0.60	$3.60	262,144	—
Novita	unknown	$0.60	$3.60	262,144	99.3%
StreamLake	unknown	$0.60	$3.60	256,000	96.2%
Venice	unknown	$0.75	$4.50	128,000	—

Buying direct from Alibaba / Qwen costs $0.60 in and $3.60 out per million. That is 91% dearer than the cheapest routed endpoint on a 3:1 blend.

Mistral Medium 3.5

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Mistral direct	first-party	$1.50	$7.50	—	—
Mistral	unknown	$1.50	$7.50	262,144	100.0%

Buying direct from Mistral costs $1.50 in and $7.50 out per million.

Kimi K2.5

10 providers · fp4, fp8, int4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Moonshot direct	first-party	$0.60	$3.00	—	—
DigitalOcean	unknown	$0.38	$2.02	262,144	98.4%
DeepInfra	fp4	$0.45	$2.25	262,144	100.0%
SiliconFlow	int4	$0.45	$2.25	262,144	100.0%
AtlasCloud	int4	$0.49	$2.50	262,144	99.8%
StreamLake	fp8	$0.54	$2.70	256,000	99.9%
Novita	unknown	$0.57	$2.85	262,144	97.5%
Amazon Bedrock	unknown	$0.60	$3.00	262,144	99.6%
Moonshot	int4	$0.60	$3.00	262,144	100.0%
Phala	unknown	$0.60	$3.00	262,144	—
Venice	unknown	$0.56	$3.50	256,000	100.0%

Buying direct from Moonshot costs $0.60 in and $3.00 out per million, cached input $0.100. That is 52% dearer than the cheapest routed endpoint on a 3:1 blend.

GLM 4.6

5 providers · bf16, fp4, fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Z.ai / Zhipu direct	first-party	$0.60	$2.20	—	—
Venice	fp4	$0.43	$1.75	198,000	100.0%
DeepInfra	fp4	$0.50	$2.00	202,752	—
Novita	bf16	$0.55	$2.20	204,800	98.5%
AtlasCloud	fp8	$0.60	$2.20	202,752	100.0%
Z.ai / Zhipu	fp4	$0.60	$2.20	202,752	99.3%

Buying direct from Z.ai / Zhipu costs $0.60 in and $2.20 out per million, cached input $0.110. That is 32% dearer than the cheapest routed endpoint on a 3:1 blend.

Qwen3.5-122B-A10B

5 providers · bf16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.40	$3.20	—	—
Alibaba / Qwen	unknown	$0.26	$2.08	262,144	100.0%
SiliconFlow	fp8	$0.26	$2.08	262,144	100.0%
DeepInfra	fp4	$0.29	$2.40	262,144	99.7%
AtlasCloud	fp8	$0.30	$2.40	262,144	99.9%
Novita	bf16	$0.40	$3.20	262,144	99.8%

Buying direct from Alibaba / Qwen costs $0.40 in and $3.20 out per million. That is 54% dearer than the cheapest routed endpoint on a 3:1 blend.

LongCat 2.0

1 providers · fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
AtlasCloud	fp8	$0.30	$1.20	1,048,756	100.0%

Not listed above: Meituan directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

GLM 4.7

9 providers · fp16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Z.ai / Zhipu direct	first-party	$0.60	$2.20	—	—
DeepInfra	fp4	$0.40	$1.75	202,752	99.7%
StreamLake	fp8	$0.48	$1.76	200,000	96.4%
AtlasCloud	fp8	$0.52	$1.85	202,752	96.7%
Novita	fp8	$0.54	$1.98	204,800	95.4%
Google	unknown	$0.60	$2.20	200,000	100.0%
Z.ai / Zhipu	fp4	$0.60	$2.20	202,752	94.2%
Mancer 2	fp4	$0.60	$2.50	131,072	98.4%
Venice	fp4	$0.55	$2.65	198,000	98.9%
Cerebras	fp16	$2.25	$2.75	131,072	100.0%

Buying direct from Z.ai / Zhipu costs $0.60 in and $2.20 out per million, cached input $0.110. That is 36% dearer than the cheapest routed endpoint on a 3:1 blend.

DeepSeek V3.2

14 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
StreamLake	fp8	$0.21	$0.32	128,000	99.4%
AtlasCloud	fp8	$0.26	$0.38	163,840	99.8%
DeepInfra	fp4	$0.26	$0.38	163,840	95.8%
SiliconFlow	fp8	$0.26	$0.42	163,840	99.2%
Novita	fp8	$0.27	$0.40	163,840	99.9%
Baidu	fp8	$0.28	$0.42	131,072	100.0%
GMICloud	fp8	$0.29	$0.43	163,840	100.0%
Venice	unknown	$0.33	$0.48	160,000	91.8%
DigitalOcean	unknown	$0.25	$0.80	163,840	98.4%
Alibaba / Qwen	fp8	$0.37	$1.11	131,072	96.9%
Friendli	unknown	$0.50	$1.50	163,840	100.0%
Google	unknown	$0.56	$1.68	163,840	100.0%
Phala	unknown	$1.00	$1.00	163,840	100.0%
SambaNova	unknown	$3.00	$4.50	32,768	—

Not listed above: DeepSeek directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

DeepSeek V3.1 Terminus

5 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp4	$0.27	$0.95	163,840	100.0%
Novita	fp8	$0.27	$1.00	131,072	100.0%
SiliconFlow	fp8	$0.27	$1.00	163,840	98.9%
AtlasCloud	fp8	$0.30	$0.95	131,072	100.0%
StreamLake	unknown	$0.34	$1.03	128,000	100.0%

Gemma 4 31B

19 providers · bf16, fp16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenInference	bf16	$0.08	$0.35	262,144	94.7%
DeepInfra	fp4	$0.09	$0.34	262,144	100.0%
CoreWeave	bf16	$0.10	$0.34	262,144	97.0%
Venice	bf16	$0.12	$0.36	256,000	96.5%
Chutes	fp4	$0.12	$0.37	131,072	8.1%
DeepInfra	fp8	$0.13	$0.38	262,144	99.1%
SiliconFlow	fp8	$0.13	$0.40	262,144	98.9%
Crusoe	unknown	$0.14	$0.40	262,144	93.9%
Friendli	unknown	$0.14	$0.40	262,144	99.7%
Morph	fp4	$0.14	$0.40	175,000	96.5%
Novita	bf16	$0.14	$0.40	262,144	93.1%
Parasail	fp8	$0.15	$0.40	262,144	87.6%
Phala	unknown	$0.15	$0.46	262,144	94.8%
DeepInfra	fp8	$0.27	$0.76	131,072	100.0%
Together	unknown	$0.28	$0.86	262,144	—
Together	unknown	$0.39	$0.97	262,144	—
SambaNova	unknown	$0.38	$1.15	262,144	98.9%
ModelRun	fp4	$0.75	$1.00	262,144	100.0%
Cerebras	fp16	$0.99	$1.49	131,072	99.9%

Not listed above: Google directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Ring-2.6-1T

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Novita	unknown	$0.07	$0.62	262,144	100.0%

Not listed above: InclusionAI directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Grok 4.3

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
xAI direct	first-party	$1.25	$2.50	—	—
xAI	unknown	$1.25	$2.50	1,000,000	99.9%
xAI	unknown	$1.25	$2.50	1,000,000	99.9%
xAI	unknown	$2.50	$5.00	1,000,000	99.9%
xAI	unknown	$2.50	$5.00	1,000,000	99.9%

Buying direct from xAI costs $1.25 in and $2.50 out per million, cached input $0.200.

Qwen3.6 35B A3B

9 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.25	$1.49	—	—
Venice	fp8	$0.10	$0.95	256,000	100.0%
DeepInfra	fp8	$0.10	$0.95	262,144	95.8%
AkashML	fp8	$0.14	$1.00	262,144	100.0%
Parasail	fp8	$0.15	$1.00	262,144	99.9%
AtlasCloud	fp8	$0.19	$1.11	262,144	100.0%
Io Net	fp8	$0.19	$1.19	262,140	100.0%
Phala	unknown	$0.20	$1.27	262,144	99.9%
CoreWeave	fp8	$0.25	$1.25	262,144	99.9%
SiliconFlow	fp8	$0.20	$1.60	262,144	99.9%

Buying direct from Alibaba / Qwen costs $0.25 in and $1.49 out per million. That is 79% dearer than the cheapest routed endpoint on a 3:1 blend.

Not listed above: Alibaba / Qwen directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

o1

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$15.00	$60.00	200,000	—

Step 3.7 Flash

3 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	unknown	$0.20	$1.15	262,144	100.0%
Novita	fp8	$0.20	$1.15	262,144	99.1%
StepFun	fp8	$0.20	$1.15	256,000	99.7%

Gemma 4 26B A4B

8 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp8	$0.07	$0.34	262,144	98.9%
Cloudflare	unknown	$0.10	$0.30	256,000	100.0%
NextBit	bf16	$0.12	$0.40	262,144	99.9%
SiliconFlow	fp8	$0.12	$0.40	262,144	99.6%
Novita	bf16	$0.13	$0.40	262,144	98.5%
Parasail	bf16	$0.13	$0.40	262,144	99.3%
Venice	bf16	$0.13	$0.40	256,000	99.0%
Google	unknown	$0.15	$0.60	262,144	98.3%

GPT-5

3 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$1.25	$10.00	400,000	—
OpenAI	unknown	$1.25	$10.00	400,000	96.6%
Azure	unknown	$1.38	$11.00	400,000	—

Qwen3.5-35B-A3B

8 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.25	$2.00	—	—
AkashML	fp8	$0.14	$1.00	262,144	100.0%
DeepInfra	fp8	$0.14	$1.00	262,144	100.0%
Parasail	fp8	$0.15	$1.00	262,144	100.0%
Alibaba / Qwen	unknown	$0.16	$1.30	262,144	99.4%
CoreWeave	fp8	$0.25	$1.25	262,144	88.4%
Venice	unknown	$0.31	$1.25	256,000	97.7%
AtlasCloud	fp8	$0.22	$1.80	262,144	100.0%
SiliconFlow	fp8	$0.24	$1.80	262,144	—

Buying direct from Alibaba / Qwen costs $0.25 in and $2.00 out per million. That is 94% dearer than the cheapest routed endpoint on a 3:1 blend.

Qwen3 Coder Next

4 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Parasail	bf16	$0.12	$0.80	262,144	100.0%
StreamLake	unknown	$0.18	$0.90	256,000	—
Novita	fp8	$0.20	$1.50	262,144	100.0%
Alibaba / Qwen	unknown	$0.30	$1.50	262,144	—

Gemini 3.1 Flash Lite

7 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$0.12	$0.75	1,048,576	98.2%
Google AI Studio	unknown	$0.12	$0.75	1,048,576	99.7%
Google	unknown	$0.25	$1.50	1,048,576	98.2%
Google AI Studio	unknown	$0.25	$1.50	1,048,576	99.7%
Google	unknown	$0.28	$1.65	1,048,576	—
Google	unknown	$0.45	$2.70	1,048,576	98.2%
Google AI Studio	unknown	$0.45	$2.70	1,048,576	99.7%

Gemini 2.5 Pro

8 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Google	unknown	$0.62	$5.00	1,048,576	99.8%
Google AI Studio	unknown	$0.62	$5.00	1,048,576	99.2%
Google	unknown	$1.25	$10.00	1,048,576	99.8%
Google	unknown	$1.25	$10.00	1,048,576	—
Google	unknown	$1.25	$10.00	1,048,576	—
Google AI Studio	unknown	$1.25	$10.00	1,048,576	99.2%
Google	unknown	$2.25	$18.00	1,048,576	99.8%
Google AI Studio	unknown	$2.25	$18.00	1,048,576	99.2%

Mercury 2

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Inception	unknown	$0.25	$0.75	128,000	100.0%

gpt-oss-120b

20 providers · bf16, fp16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
CoreWeave	fp4	$0.03	$0.17	131,072	94.2%
DeepInfra	bf16	$0.04	$0.17	131,072	99.9%
Novita	fp4	$0.05	$0.25	131,072	83.1%
DigitalOcean	unknown	$0.06	$0.39	128,000	100.0%
SiliconFlow	fp8	$0.05	$0.45	131,072	99.4%
AkashML	bf16	$0.04	$0.49	131,072	99.9%
Google	unknown	$0.09	$0.36	131,072	92.9%
Mancer 2	fp8	$0.08	$0.50	131,072	99.3%
BaseTen	fp4	$0.10	$0.50	128,072	100.0%
Amazon Bedrock	unknown	$0.15	$0.60	131,072	100.0%
Amazon Bedrock	unknown	$0.15	$0.60	131,072	—
DeepInfra	bf16	$0.15	$0.60	131,072	99.5%
Groq	unknown	$0.15	$0.60	131,072	100.0%
Nebius	fp4	$0.15	$0.60	131,072	99.9%
Phala	unknown	$0.15	$0.60	131,072	100.0%
Together	unknown	$0.15	$0.60	131,072	94.9%
Parasail	fp4	$0.10	$0.75	131,072	100.0%
Mara	unknown	$0.15	$0.75	131,072	86.2%
SambaNova	unknown	$0.14	$0.95	131,072	92.5%
Cerebras	fp16	$0.35	$0.75	131,072	99.8%

Not listed above: OpenAI directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Qwen3.5-9B

5 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	bf16	$0.10	$0.15	262,144	99.9%
SiliconFlow	fp8	$0.10	$0.15	262,144	69.4%
Venice	fp8	$0.10	$0.15	256,000	100.0%
Parasail	bf16	$0.10	$0.25	262,144	100.0%
Together	unknown	$0.17	$0.25	262,144	99.6%

Command A

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Cohere	unknown	$2.50	$10.00	256,000	100.0%

Mistral Small 4

2 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Mistral direct	first-party	$0.15	$0.60	—	—
Mistral	unknown	$0.15	$0.60	262,144	100.0%
Venice	fp8	$0.19	$0.75	256,000	100.0%

Buying direct from Mistral costs $0.15 in and $0.60 out per million.

Trinity Large Thinking

2 providers · fp4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Parasail	fp4	$0.22	$0.85	262,144	100.0%
Arcee AI	unknown	$0.25	$0.80	262,144	—

Not listed above: Arcee Ai directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Ling-3.0-flash

2 providers · bf16, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Novita	unknown	$0.02	$0.06	262,144	100.0%
DeepInfra	bf16	$0.06	$0.18	131,072	99.6%

Muse Spark 1.2

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Meta	unknown	$1.25	$4.25	1,048,576	100.0%

Qwen3.8 Max

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen	unknown	$2.00	$6.00	1,000,000	100.0%

GPT-4o

2 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$2.50	$10.00	128,000	100.0%
OpenAI	unknown	$2.50	$10.00	128,000	99.8%

DeepSeek V3

3 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepSeek direct	first-party	$0.14	$0.28	—	—
StreamLake	unknown	$0.26	$1.03	128,000	99.8%
DeepInfra	fp4	$0.32	$0.89	163,840	99.2%
Novita	fp8	$0.40	$1.30	64,000	100.0%

Buying direct from DeepSeek costs $0.14 in and $0.28 out per million, cached input $0.003. That is 61% cheaper than the cheapest routed endpoint on a 3:1 blend.

GPT-4 Turbo

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$10.00	$30.00	128,000	100.0%

DeepSeek V3 0324

4 providers · bf16, fp4, fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp4	$0.24	$0.90	163,840	97.8%
SiliconFlow	fp8	$0.25	$1.00	163,840	99.6%
Novita	fp8	$0.27	$1.12	163,840	99.9%
Crusoe	bf16	$0.50	$1.50	163,840	100.0%

gpt-oss-20b

12 providers · bf16, fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
CoreWeave	fp4	$0.03	$0.13	131,072	99.6%
DeepInfra	bf16	$0.03	$0.14	131,072	100.0%
Parasail	fp4	$0.03	$0.15	131,072	99.3%
Novita	fp4	$0.04	$0.15	131,072	100.0%
Phala	unknown	$0.04	$0.15	131,072	96.6%
SiliconFlow	fp8	$0.04	$0.18	131,072	95.3%
Together	unknown	$0.05	$0.20	131,072	—
Amazon Bedrock	unknown	$0.07	$0.15	131,072	—
Amazon Bedrock	unknown	$0.07	$0.15	131,072	97.5%
Google	unknown	$0.07	$0.25	131,072	98.4%
Fireworks	unknown	$0.07	$0.30	131,072	98.9%
Groq	unknown	$0.07	$0.30	131,072	99.9%

Not listed above: OpenAI directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Mistral Medium 3.1

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Mistral	unknown	$0.40	$2.00	131,072	100.0%

GPT-4.1 Mini

3 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$0.40	$1.60	1,047,576	100.0%
OpenAI	unknown	$0.40	$1.60	1,047,576	99.6%
Azure	unknown	$0.44	$1.76	1,047,576	100.0%

Llama 4 Maverick

5 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DigitalOcean	unknown	$0.20	$0.70	128,000	100.0%
DeepInfra	fp8	$0.20	$0.80	1,048,576	99.6%
Novita	fp8	$0.27	$0.85	1,048,576	99.0%
Parasail	fp8	$0.35	$1.00	524,288	99.9%
Google	unknown	$0.35	$1.15	524,288	98.0%

Not listed above: Meta directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

o3 Mini

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$1.10	$4.40	200,000	100.0%

Solar Pro 3

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Upstage	unknown	$0.15	$0.60	131,072	—

GPT-5 Mini

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.12	$1.00	400,000	99.8%
Azure	unknown	$0.25	$2.00	400,000	100.0%
OpenAI	unknown	$0.25	$2.00	400,000	99.8%
Azure	unknown	$0.28	$2.20	400,000	—

Qwen3 32B

4 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.70	$2.80	—	—
DeepInfra	fp8	$0.08	$0.28	40,960	100.0%
Nebius	fp8	$0.10	$0.30	40,960	99.8%
SiliconFlow	fp8	$0.14	$0.57	131,072	100.0%
Groq	unknown	$0.29	$0.59	131,072	100.0%

Buying direct from Alibaba / Qwen costs $0.70 in and $2.80 out per million. That is 842% dearer than the cheapest routed endpoint on a 3:1 blend.

Qwen3 14B

3 providers · fp8, int4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.35	$1.40	—	—
NextBit	int4	$0.10	$0.22	40,960	99.3%
DeepInfra	fp8	$0.12	$0.24	40,960	99.9%
Alibaba / Qwen	unknown	$0.23	$0.91	131,072	100.0%

Buying direct from Alibaba / Qwen costs $0.35 in and $1.40 out per million. That is 371% dearer than the cheapest routed endpoint on a 3:1 blend.

GPT-4

2 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$30.00	$60.00	8,191	100.0%
OpenAI	unknown	$30.00	$60.00	8,191	100.0%

GPT-4o-mini

3 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$0.15	$0.60	128,000	100.0%
OpenAI	unknown	$0.15	$0.60	128,000	100.0%
Azure	unknown	$0.17	$0.66	128,000	—

GPT-4.1 Nano

3 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Azure	unknown	$0.10	$0.40	1,047,576	99.9%
OpenAI	unknown	$0.10	$0.40	1,047,576	99.5%
Azure	unknown	$0.11	$0.44	1,047,576	—

GPT-3.5 Turbo

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
OpenAI	unknown	$0.50	$1.50	16,385	100.0%

Granite 4.1 8B

1 providers · bf16

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
CoreWeave	bf16	$0.05	$0.10	131,072	100.0%

Not listed above: IBM directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Qwen3 8B

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.18	$0.70	—	—
Alibaba / Qwen	unknown	$0.12	$0.45	131,072	100.0%

Buying direct from Alibaba / Qwen costs $0.18 in and $0.70 out per million. That is 54% dearer than the cheapest routed endpoint on a 3:1 blend.

Llama 4 Scout

4 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp8	$0.10	$0.30	327,680	100.0%
Groq	unknown	$0.11	$0.34	131,072	99.9%
Novita	bf16	$0.18	$0.59	131,072	99.9%
Google	unknown	$0.25	$0.70	1,310,720	99.7%

Not listed above: Meta directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Grok Build 0.1

4 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
xAI direct	first-party	$1.00	$2.00	—	—
xAI	unknown	$1.00	$2.00	256,000	100.0%
xAI	unknown	$1.00	$2.00	256,000	—
xAI	unknown	$2.00	$4.00	256,000	100.0%
xAI	unknown	$2.00	$4.00	256,000	—

Buying direct from xAI costs $1.00 in and $2.00 out per million, cached input $0.200.

Nemotron 3 Ultra 550B

4 providers · fp4, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Nvidia direct	first-party	$0.50	$2.50	—	—
DeepInfra	fp4	$0.50	$2.20	262,144	91.2%
BaseTen	fp4	$0.60	$2.40	202,800	98.8%
Venice	fp8	$0.62	$3.12	256,000	87.4%
Together	unknown	$0.60	$3.60	512,288	98.2%

Buying direct from Nvidia costs $0.50 in and $2.50 out per million, cached input $0.150. That is 8% dearer than the cheapest routed endpoint on a 3:1 blend.

Not listed above: Nvidia directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Nemotron 3 Super 120B

3 providers · bf16, fp4, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Nvidia direct	first-party	$0.20	$0.80	—	—
DeepInfra	bf16	$0.08	$0.40	262,144	92.8%
DigitalOcean	unknown	$0.17	$0.36	1,000,000	92.9%
Nebius	fp4	$0.30	$0.90	8,000	98.0%

Buying direct from Nvidia costs $0.20 in and $0.80 out per million. That is 114% dearer than the cheapest routed endpoint on a 3:1 blend.

Not listed above: Nvidia directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

North Mini Code

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Cohere direct	first-party	$0.00	$0.00	—	—
Cohere	unknown	$0.00	$0.00	256,000	96.3%

Buying direct from Cohere costs $0.00 in and $0.00 out per million.

Mistral Small 3.1

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Cloudflare	unknown	$0.35	$0.55	128,000	100.0%

Not listed above: Mistral directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

DeepSeek R1

1 providers · fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Novita	fp8	$0.70	$2.50	64,000	100.0%

Qwen3 235B A22B 2507

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Alibaba / Qwen direct	first-party	$0.70	$2.80	—	—
Alibaba / Qwen	unknown	$0.45	$1.82	131,072	100.0%

Buying direct from Alibaba / Qwen costs $0.70 in and $2.80 out per million. That is 54% dearer than the cheapest routed endpoint on a 3:1 blend.

Mistral Large 3

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Mistral direct	first-party	$0.50	$1.50	—	—
Mistral	unknown	$2.00	$6.00	128,000	100.0%

Buying direct from Mistral costs $0.50 in and $1.50 out per million. That is 75% cheaper than the cheapest routed endpoint on a 3:1 blend.

Nemotron 3 Nano 30B

3 providers · fp4, fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Nvidia direct	first-party	$0.00	$0.00	—	—
Crusoe	fp8	$0.05	$0.20	262,144	98.2%
DeepInfra	fp4	$0.05	$0.20	262,144	96.3%
Novita	fp4	$0.05	$0.20	262,144	0.0%

Buying direct from Nvidia costs $0.00 in and $0.00 out per million.

Not listed above: Nvidia directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Nemotron 3 Nano Omni 30B

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Nvidia direct	first-party	$0.00	$0.00	—	—
Nvidia	unknown	$0.00	$0.00	256,000	99.2%

Buying direct from Nvidia costs $0.00 in and $0.00 out per million.

Mistral Small 3.2

3 providers · bf16, fp8

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Mistral direct	first-party	$0.10	$0.30	—	—
DeepInfra	fp8	$0.07	$0.20	128,000	99.8%
Venice	fp8	$0.09	$0.25	256,000	99.9%
Parasail	bf16	$0.09	$0.30	131,072	98.8%

Buying direct from Mistral costs $0.10 in and $0.30 out per million. That is 41% dearer than the cheapest routed endpoint on a 3:1 blend.

Not listed above: Mistral directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Qwen3 30B A3B 2507

2 providers · fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp8	$0.12	$0.50	40,960	100.0%
Alibaba / Qwen	unknown	$0.13	$0.52	131,072	100.0%

Gemma 3 27B

5 providers · bf16, fp8, unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	fp8	$0.08	$0.16	131,072	99.8%
Novita	bf16	$0.12	$0.20	98,304	99.3%
Nebius	fp8	$0.10	$0.30	110,000	99.4%
Parasail	fp8	$0.08	$0.45	131,072	99.9%
Phala	unknown	$0.15	$0.46	262,144	89.2%

Not listed above: Google directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Gemma 3 12B

1 providers · bf16

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	bf16	$0.05	$0.15	131,072	100.0%

Not listed above: Google directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Gemma 3n E4B

1 providers · unknown

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
Together	unknown	$0.06	$0.12	32,768	100.0%

Not listed above: Google directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

Gemma 3 4B

1 providers · bf16

Served by	Quant	$/M in	$/M out	Context	Uptime 30m
DeepInfra	bf16	$0.05	$0.10	131,072	100.0%

Not listed above: Google directly. The lab sells this model from its own API but does not route it through the aggregator these figures come from, so its price and precision cannot be compared here.

05 CostReal cost Best value What a plan buys

What it actually costs

Observed, not list price

Aggregates across opencode’s subscriber base: real cost per coding session after caching. The spread is far wider than list prices suggest, and the cache ratio explains much of it.

Model	$ / session	Effective $/M	Cache ratio	Tokens / session	Sessions observed
Xiaomi MiMo-V2.5	$0.0722	$0.01	94.3%	5.9M	1,448,024
DeepSeek V4 Flash	$0.1393	$0.01	96.4%	14.6M	12,096,796
Tencent Hy3	$0.2951	$0.16	87.0%	1.8M	977
Xiaomi MiMo-V2.5-Pro	$0.3802	$0.06	95.5%	5.9M	373,253
DeepSeek V4 Pro	$0.4384	$0.05	96.8%	8.1M	3,949,114
Qwen3.7 Plus	$0.5280	$0.17	94.6%	3.0M	645,425
MiniMax-M3	$0.5434	$0.08	94.3%	7.1M	1,245,848
Kimi K2.6	$0.8180	$0.28	91.2%	2.9M	314,084
Grok 4.5	$0.8340	$0.67	86.7%	1.2M	79,708
Kimi K2.7 Code	$1.0948	$0.25	94.2%	4.3M	720,175
Qwen3.7 Max	$1.5594	$0.96	95.1%	1.6M	228,040
GLM-5.1	$1.9516	$0.61	75.1%	3.2M	127,184
Kimi K3	$2.0958	$0.69	92.0%	3.0M	426,856
GLM-5.2	$2.5679	$0.51	80.6%	5.0M	1,750,878

Best value

Two ways to pay, two different answers

Value means something different depending on how you are billed, so this splits in two. On a subscription what matters is how much quality-weighted work a fixed monthly fee yields; on metered billing it is quality per unit of token spend. The same model can look strong under one and ordinary under the other.

On a subscription

Coding index times sessions per month, divided by what the plan costs. It answers a narrower question than it looks — how much quality-weighted work a subscription yields, assuming the cheap model is good enough for the job in hand.

#	Plan	Model	$/mo	Sessions/mo	Coding index	Value: index x sessions per $
1	opencode Go	Xiaomi MiMo-V2.5	$10	831	56.8	██████████████████████████████████████████████ 4,720
2	opencode Go	DeepSeek V4 Flash	$10	431	69.1	█████████████████████████████ 2,976
3	Qoder Pro	Xiaomi MiMo-V2.5	$20	554	56.8	███████████████ 1,573
4	GitHub Max	Xiaomi MiMo-V2.5	$100	2,770	56.8	███████████████ 1,573
5	Cursor Ultra	Xiaomi MiMo-V2.5	$200	5,540	56.8	███████████████ 1,573
6	Qoder Pro+	Xiaomi MiMo-V2.5	$60	1,662	56.8	███████████████ 1,573
7	Qoder Ultra	Xiaomi MiMo-V2.5	$200	5,540	56.8	███████████████ 1,573
8	GitHub Pro+	Xiaomi MiMo-V2.5	$39	970	56.8	██████████████ 1,412
9	opencode Go	Tencent Hy3	$10	203	58.8	████████████ 1,196
10	GitHub Pro	Xiaomi MiMo-V2.5	$10	208	56.8	████████████ 1,180
11	Qoder Pro	DeepSeek V4 Flash	$20	287	69.1	██████████ 992
12	Qoder Pro+	DeepSeek V4 Flash	$60	861	69.1	██████████ 992
13	GitHub Max	DeepSeek V4 Flash	$100	1,436	69.1	██████████ 992
14	Cursor Ultra	DeepSeek V4 Flash	$200	2,872	69.1	██████████ 992

Absence from this table is not a verdict. Only plans that publish a dollar budget can be ranked here, which is three vendors out of fourteen. The rest are excluded for being unmeasurable, not for being poor value — and several of them are deliberately generous. MiniMax is the clearest case: it publishes no quota at all, yet the figure reverse-engineered further down this page (about 32B tokens a month on the $120 tier) works out at roughly 4,200 sessions, which would place it third here — ahead of every GitHub and Cursor tier listed. Z.ai, Alibaba and Kimi meter in prompts or requests with no published token cost, so they cannot be placed at all. Read the table as the best of what can be measured, never as the best available.

Read it with the leaderboard beside you, too. A weak model on a generous plan can top this table while still being the wrong tool: the metric rewards volume, and quality enters only linearly. It is a guide to where the money goes furthest, not to what you should run for work that has to be right first time.

Pay per token

Coding index per dollar per million tokens, blended 3:1. The list column ranks what you are quoted; the observed column uses the effective rate actually paid in production once caching is working, which is the number that decides a monthly bill.

#	Model	Coding index	$/M list	Value on list	$/M observed	Value observed
1	Ling-3.0-flash · InclusionAI	50.6	$0.032	1,606	—	—
2	Solar Pro 4 · Upstage	52.7	$0.052	1,004	—	—
3	gpt-oss-120b · OpenAI	30.4	$0.065	468	—	—
4	DeepSeek V4 Flash · DeepSeek	69.1	$0.175	395	$0.01	6,910
5	gpt-oss-20b · OpenAI	20.7	$0.055	376	—	—
6	Xiaomi MiMo-V2.5 · Xiaomi	56.8	$0.175	325	$0.01	5,680
7	GPT-5.6 Luna · OpenAI	71.4	$0.225	317	—	—
8	Gemma 4 31B · Google	43.4	$0.160	271	—	—
9	Qwen3.5-9B · Alibaba / Qwen	28.7	$0.112	255	—	—
10	Tencent Hy3 · Tencent	58.8	$0.231	255	$0.16	368
11	Gemma 4 26B A4B · Google	39.3	$0.190	207	—	—
12	Ring-2.6-1T · InclusionAI	42.8	$0.212	201	—	—
13	Nemotron 3.5 Lightning · Nvidia	26.8	$0.138	195	—	—
14	Nemotron 3 Nano 30B · Nvidia	14.4	$0.087	165	—	—
15	Granite 4.1 8B · IBM	9.5	$0.062	152	—	—
16	DeepSeek V3.2 · DeepSeek	44.2	$0.302	146	—	—
17	Nex-N2-Pro · Nex AGI	59.1	$0.438	135	—	—
18	DeepSeek V4 Pro · DeepSeek	68.8	$0.544	127	$0.05	1,376
19	Qwen3 Coder Next · Alibaba / Qwen	36.2	$0.290	125	—	—
20	GPT-5.4 Nano · OpenAI	56.1	$0.463	121	—	—
21	Qwen3 32B · Alibaba / Qwen	15.3	$0.130	118	—	—
22	Qwen3.6 35B A3B · Alibaba / Qwen	41.9	$0.362	116	—	—
23	KAT-Coder-Pro V2 · Kwaipilot	59.5	$0.525	113	—	—
24	MiniMax-M3 · MiniMax	58.6	$0.525	112	$0.08	732
25	Xiaomi MiMo-V2.5-Pro · Xiaomi	60.2	$0.544	111	$0.06	1,003
26	Mistral Small 4 · Mistral	26.6	$0.262	101	—	—
27	MiniMax M2.7 · MiniMax	52.6	$0.525	100	—	—
28	Qwen3.7 Plus · Alibaba / Qwen	55.9	$0.560	100	$0.17	329
29	DeepSeek V3.1 Terminus · DeepSeek	43.5	$0.440	99	—	—
30	Mistral Small 3.2 · Mistral	12.5	$0.133	94	—	—
31	Qwen3 14B · Alibaba / Qwen	13.8	$0.150	92	—	—
32	Step 3.7 Flash · StepFun	39.6	$0.438	91	—	—
33	LongCat 2.0 · Meituan	45.3	$0.525	86	—	—
34	Mercury 2 · Inception	31.1	$0.375	83	—	—
35	Gemma 3 12B · Google	5.8	$0.075	77	—	—
36	Muse Glimmer · Meta	49.0	$0.637	77	—	—
37	Qwen3.6 Plus · Alibaba / Qwen	54.5	$0.731	75	—	—
38	Trinity Large Thinking · Arcee Ai	25.8	$0.378	68	—	—
39	Mistral Small 3.1 · Mistral	26.3	$0.402	65	—	—
40	GPT-4.1 Nano · OpenAI	11.1	$0.175	63	—	—
41	Solar Pro 3 · Upstage	16.2	$0.262	62	—	—
42	Gemini 3.1 Flash Lite · Google	34.7	$0.562	62	—	—
43	GLM 4.7 · Z.ai / Zhipu	45.3	$0.738	61	—	—
44	Kimi K2.6 · Moonshot	61.8	$1.010	61	$0.28	221
45	Qwen3.5-35B-A3B · Alibaba / Qwen	37.0	$0.619	60	—	—
46	Gemma 3 27B · Google	10.1	$0.172	59	—	—
47	Gemini 3.5 Flash-Lite · Google	49.3	$0.850	58	—	—
48	Qwen3 30B A3B 2507 · Alibaba / Qwen	12.1	$0.215	56	—	—
49	Qwen3.5-122B-A10B · Alibaba / Qwen	45.7	$0.817	56	—	—
50	Qwen3.5 397B A17B · Alibaba / Qwen	48.2	$0.877	55	—	—
51	Llama 4 Scout · Meta	8.2	$0.150	55	—	—
52	DeepSeek V3 · DeepSeek	23.0	$0.450	51	—	—
53	GLM 4.6 · Z.ai / Zhipu	45.8	$0.963	48	—	—
54	Llama 4 Maverick · Meta	16.3	$0.350	47	—	—
55	Gemini 3.6 Flash · Google	69.2	$1.500	46	—	—
56	Qwen3 8B · Alibaba / Qwen	9.0	$0.202	45	—	—
57	DeepSeek V3 0324 · DeepSeek	21.2	$0.483	44	—	—
58	GPT-4o-mini · OpenAI	11.4	$0.262	43	—	—
59	Gemma 3 4B · Google	2.7	$0.062	43	—	—
60	Kimi K2.7 Code · Moonshot	60.8	$1.407	43	$0.25	243
61	Gemma 3n E4B · Google	3.2	$0.075	43	—	—
62	Grok Build 0.1 · xAI	51.5	$1.250	41	—	—
63	Kimi K2.5 · Moonshot	46.8	$1.140	41	—	—
64	Qwen3.6 27B · Alibaba / Qwen	53.7	$1.350	40	—	—
65	GLM-5.2 · Z.ai / Zhipu	68.8	$1.828	38	$0.51	135
66	Nemotron 3 Ultra 550B · Nvidia	49.3	$1.350	37	—	—
67	Muse Spark 1.2 · Meta	72.2	$2.000	36	—	—
68	Muse Spark 1.1 · Meta	71.3	$2.000	36	—	—
69	GPT-5.6 Terra · OpenAI	76.7	$2.250	34	—	—
70	GPT-5.4 Mini · OpenAI	56.1	$1.688	33	—	—
71	Inkling · Thinking Machines	52.1	$1.725	30	—	—
72	Qwen3.7 Max · Alibaba / Qwen	66.0	$2.213	30	$0.96	69
73	GPT-4.1 Mini · OpenAI	20.2	$0.700	29	—	—
74	Qwen3 235B A22B 2507 · Alibaba / Qwen	22.1	$0.796	28	—	—
75	Grok 4.3 · xAI	42.2	$1.562	27	—	—
76	GLM-5.1 · Z.ai / Zhipu	55.8	$2.150	26	$0.61	91
77	Mistral Medium 3.1 · Mistral	20.5	$0.800	26	—	—
78	Grok 4.6 · xAI	76.8	$3.000	26	—	—
79	Grok 4.5 · xAI	72.4	$3.000	24	$0.67	108
80	Qwen3.8 Max · Alibaba / Qwen	71.8	$3.000	24	—	—
81	GPT-5 Mini · OpenAI	15.6	$0.688	23	—	—
82	DeepSeek R1 · DeepSeek	24.6	$1.150	21	—	—
83	Gemini 3.5 Flash · Google	70.1	$3.375	21	—	—
84	Claude Sonnet 5 · Anthropic	71.5	$4.000	18	—	—
85	Mistral Medium 3.5 · Mistral	46.9	$3.000	16	—	—
86	Gemini 3.1 Pro Preview · Google	68.8	$4.500	15	—	—
87	GPT-5.1 · OpenAI	49.4	$3.438	14	—	—
88	GPT-3.5 Turbo · OpenAI	10.7	$0.750	14	—	—
89	Kimi K3 · Moonshot	76.2	$6.000	13	$0.69	110
90	GPT-5.4 · OpenAI	71.1	$5.625	13	—	—
91	GPT-5 · OpenAI	37.8	$3.438	11	—	—
92	Claude Sonnet 4.6 · Anthropic	63.0	$6.000	10	—	—
93	Gemini 2.5 Pro · Google	33.3	$3.438	10	—	—
94	o3 Mini · OpenAI	16.3	$1.925	8	—	—
95	Claude Opus 5 · Anthropic	78.0	$10.000	8	—	—
96	Claude Opus 4.8 · Anthropic	74.3	$10.000	7	—	—
97	Claude Opus 4.7 · Anthropic	73.6	$10.000	7	—	—
98	GPT-5.6 Sol · OpenAI	78.3	$11.250	7	—	—
99	Mistral Large 3 · Mistral	20.1	$3.000	7	—	—
100	GPT-5.5 · OpenAI	74.9	$11.250	7	—	—
101	Command A · Cohere	27.8	$4.375	6	—	—
102	GPT-4o · OpenAI	24.2	$4.375	6	—	—
103	Claude Fable 5 · Anthropic	76.5	$20.000	4	—	—
104	o1 · OpenAI	39.7	$26.250	2	—	—
105	GPT-4 Turbo · OpenAI	21.5	$15.000	1	—	—
106	GPT-4 · OpenAI	13.1	$37.500	0	—	—

The two columns do not agree, and that disagreement is the point: list price is a quote, the observed rate is a bill. A model with a poor cache ratio slides down the observed column even when its quoted price looked competitive. Observed rates come from opencode's subscriber base, so models it does not serve show a dash rather than a guess.

What a plan actually buys

Included usage divided by observed cost per session. Only dollar-denominated plans can be answered this precisely — plans metered in prompts, or with no published quota, cannot appear here at all.

Plan	$/mo	Included	Xiaomi MiMo-V2.5	DeepSeek V4 Flash	Tencent Hy3	Xiaomi MiMo-V2.5-Pro	DeepSeek V4 Pro	Qwen3.7 Plus	MiniMax-M3	Kimi K2.6	Grok 4.5	Kimi K2.7 Code	Qwen3.7 Max	GLM-5.1	Kimi K3	GLM-5.2
opencode Go	$10	$60	831	430	203	157	136	113	110	73	71	54	38	30	28	23
GitHub Max	$100	$200	2,770	1,435	677	526	456	378	368	244	239	182	128	102	95	77
Cursor Ultra	$200	$400	5,540	2,871	1,355	1,052	912	757	736	488	479	365	256	204	190	155
Qoder Pro	$20	$40	554	287	135	105	91	75	73	48	47	36	25	20	19	15
Qoder Pro+	$60	$120	1,662	861	406	315	273	227	220	146	143	109	76	61	57	46
Qoder Ultra	$200	$400	5,540	2,871	1,355	1,052	912	757	736	488	479	365	256	204	190	155
GitHub Pro+	$39	$70	969	502	237	184	159	132	128	85	83	63	44	35	33	27
GitHub Pro	$10	$15	207	107	50	39	34	28	27	18	17	13	9	7	7	5
Cursor Pro+	$60	$70	969	502	237	184	159	132	128	85	83	63	44	35	33	27
Venice Max	$200	$225	3,116	1,615	762	591	513	426	414	275	269	205	144	115	107	87
Venice Pro Plus	$68	$75	1,038	538	254	197	171	142	138	91	89	68	48	38	35	29
Cursor Pro	$20	$20	277	143	67	52	45	37	36	24	23	18	12	10	9	7
Augment Code Business	$100	$100	1,385	717	338	263	228	189	184	122	119	91	64	51	47	38
Venice Pro	$18	$1	13	7	3	2	2	1	1	1	1	0	0	0	0	0

Sessions per month. Heavier models burn a budget far faster; real mileage varies with session length and cache behaviour.

06 PlansSubscriptions Multipliers Hidden quota

Subscription plans

Grouped by how verifiable the quota actually is

Most vendors do not publish a convertible quota. Effective $/token is computed only where a vendor publishes enough to do it honestly; everything else is left blank rather than guessed.

Provider	Plan	$/month	Quota disclosure	See what's left	Quota converts to units?	Checked
Xiaomi	MiMo Token Plan (Lite / Standard / Pro / Max)	$6.00 / $16.00 / $50.00 / $100.00	Convertible token quota	console only	converts	2026-08-13
opencode	opencode Go	$10.00	Dollar usage budget	console only	converts	2026-08-13
GitHub	Copilot (Pro / Pro+ / Max)	$0.00 / $10.00 / $39.00 / $100.00	Dollar usage budget	console only	converts	2026-08-13
Cursor	Pro / Pro+ / Ultra	$20.00 / $60.00 / $200.00	Dollar usage budget	console only	converts	2026-08-13
Alibaba / Qwen	Coding Plan (Pro)	$50.00	Requests / prompts	console only	partial	2026-08-13
Alibaba / Qwen	Token Plan (Individual + Team)	$6.00 / $18.00 / $68.00	No numeric quota	console only	unstated	2026-08-13
Z.ai / Zhipu	GLM Coding (Lite / Pro / Max)	$18.00 / $72.00 / $160.00	Requests / prompts	console only	partial	2026-08-14
MiniMax	Token Plan (Plus / Max / Ultra)	$20.00 / $50.00 / $120.00	No numeric quota	API	unstated	2026-08-13
Anthropic	Claude Pro / Max 5x / Max 20x	$20.00 / $100.00 / $200.00	No numeric quota	console only	unstated	2026-08-13
OpenAI	ChatGPT Plus / Pro 5x / Pro 20x	$0.00 / $8.00 / $20.00 / $100.00 / $200.00	No numeric quota	console only	partial	2026-08-13
Google	AI Plus / Pro / Ultra	$4.99 / $19.99 / $99.99 / $200.00	No numeric quota	console only	unstated	2026-08-13
Moonshot / Kimi	Moderato / Allegretto / Allegro / Vivace	$19.00 / $39.00 / $99.00 / $199.00	No numeric quota	API	partial	2026-08-13
Mistral	Le Chat Free / Pro / Team (Mistral Vibe bundled)	$0.00 / $14.99 / $24.99	No numeric quota	nowhere	unstated	2026-08-13
Windsurf (Cognition/Devin)	Pro / Max / Teams	$0.00 / $20.00 / $200.00	No numeric quota	console only	unstated	2026-08-13
Ollama	Ollama Cloud (Free / Pro / Max)	$0.00 / $20.00 / $100.00	No numeric quota	console only	unstated	2026-08-13
Venice	Venice (Free / Pro / Pro Plus / Max)	$0.00 / $18.00 / $68.00 / $200.00	Dollar usage budget	console only	converts	2026-08-13
AtlasCloud	Coding Plan (Starter / Lite / Plus / Max / Ultra / Enterprise)	$10.00 / $20.00 / $50.00 / $100.00 / $200.00 / $500.00	Requests / prompts	console only	unstated	2026-08-13
Io Net	IO Intelligence (Standard / Professional / Developer)	$0.00 / $15.00 / $150.00	Requests / prompts	console only	unstated	2026-08-13
Chutes	Plus / Pro	$10.00 / $20.00	No numeric quota	console only	unstated	2026-08-13
Morph	Usage-based (Free / Pay-per-token / Scale flat-rate)	$0.00 / $200.00	Dollar usage budget	console only	converts	2026-08-13
Qoder	Qoder (Free / Pro / Pro+ / Ultra)	$0.00 / $20.00 / $60.00 / $200.00	Dollar usage budget	console only	partial	2026-08-13
Agnes AI	Token Plan (Starter / Plus / Pro)	$4.00 / $10.00 / $50.00	Requests / prompts	console only	partial	2026-08-13
Synthetic	Synthetic (Standard)	$30.00	Dollar usage budget	console only	partial	2026-08-13
Augment Code	Augment (Business / Enterprise) — Auggie CLI + Cosmos	$100.00	Dollar usage budget	console only	converts	2026-08-13

One plan meters something else entirely. Ollama Cloud bills on “actual utilization of Ollama’s cloud infrastructure — primarily GPU time”, the only compute-metered subscription here. That inverts an assumption the rest of this page relies on: everywhere else a token costs the same whoever produced it, so model speed only affects your patience. On GPU time, speed is the quota — a model running at half the tokens per second burns roughly twice the allowance for identical output. Check the leaderboard's tok/s column before picking a model to run there.

These two columns are inverted. The vendors that let you read your quota programmatically do not tell you what you bought, and the ones that publish proper unit maths only give you a web page. MiniMax and Kimi expose an endpoint for a quota with no published units at all; Xiaomi, GitHub and opencode publish exact conversions but no API. Nobody does both. DeepSeek is the only one coherent on both axes, and only because it sells no subscription — there, remaining quota is simply money.

Only three providers document a remaining-quota endpoint, and all three are Chinese. Western billing APIs return consumed usage and never an allowance, which is why every community usage tool hardcodes the limit and does the subtraction itself. Endpoints for MiniMax, Z.ai, Kimi and DeepSeek were called directly against live keys for this table; provider coverage cross-checked against quota-sentinel.

What moves the price inside a plan

Bigger than the gap between plans

Choosing a subscription is the decision buyers agonise over, and it is usually the smaller one. Once inside a plan, what you pay per unit of work varies by model, by tier, by how you use the context and, at two vendors, by the time of day. At Qoder those axes compound into a 27-fold swing in what a credit is worth, far more than separates the plans themselves.

Vendor	Varies by	Range	Detail
Qoder	model	0.1x - 0.6x	A published credit multiplier per model. They do not track open-market price: one unit of credit buys $5.60 of Qwen3.7-Plus but only $0.68 of GLM-5.2.
Qoder	clock	down to 0.01x	Selected models are discounted 14:00-00:00 UTC, weekends and holidays included. Qwen3.7-Max falls 0.5x to 0.1x; a promoted preview model reaches 0.01x. Credit price only -- model quality is explicitly unaffected.
Qoder	tier	free - 1.6x	Lite costs nothing, Efficient 0.3x, Auto 1.0x, Performance 1.1x, Ultimate 1.6x (currently halved to 0.8x). Sits on top of the model choice, so the two compound.
Z.ai / Zhipu	clock	2x - 3x	The mirror image: quota burns 3x during peak hours and 2x off-peak, rather than being discounted. A promotion holds off-peak at 1x until 30 September 2026, after which the usable plan roughly halves at an unchanged price.
Cursor	mechanics	1.25x - 2x	Cache writes cost 1.25x input, and long context doubles above 200K tokens. Charges that depend on how you use the model rather than which one you pick.
Xiaomi	cache	50x - 120x	The largest spread found anywhere, in the subscriber's favour: a cache hit costs 2 credits per token against 100 on mimo-v2.5, and 2.5 against 300 on the Pro model. Consumption also drops to 0.8x between 00:00 and 08:00 Beijing time.
opencode	model	2x	Kimi K3 draws twice the usage of every other model in the Go plan.
Anthropic / Google / OpenAI	tier	2x - 20x	Sold as more usage rather than as a multiplier on consumption, and not comparable between vendors: Google's 4x is measured against its free tier while its 5x and 20x are against Pro, and Anthropic's multipliers are per session, not per month.
DeepSeek	clock	up to 12x	From 2026-08-16 16:00 UTC DeepSeek moves to peak/off-peak API billing with item-specific multipliers, a sharp escalation of the flat peak surcharge it had run since mid-2026. Against the current flat price, V4 Pro cache-hit input rises 6x off-peak and 12x at peak ($0.0036 -> $0.022 / $0.044 per M), cache-miss input 1.5x / 3x ($0.435 -> $0.66 / $1.32) and output 2.25x / 4.5x ($0.87 -> $1.98 / $3.96); V4 Flash scales the same way ($0.14 miss -> $0.22 / $0.44, $0.28 output -> $0.66 / $1.32). Peak is 01:00-04:00 and 06:00-10:00 UTC (Beijing working hours 09:00-12:00 and 14:00-18:00), off-peak is everything else at half the peak rate. The cache-hit multiplier is the one that bites: a coding agent replays a large cached prefix every turn at ~92-95% cache-hit, so its effective bill tracks the 12x line, not the 3x cache-miss one. Confirmed on the official pricing page 2026-08-14; the flat rate a comparison table quotes is only the pre-Aug-16 number.

Two deserve naming. Qoder and Z.ai run the same mechanism in opposite directions, one discounting the quiet hours and the other surcharging the busy ones, and only a discount looks like a discount. And Xiaomi's cache multiplier is the one that runs in the customer's favour, at up to 120x, which is why cache behaviour decides more of a coding agent's bill than the choice of model does.

Reverse-engineering a hidden quota

Derived, not official

MiniMax publishes no token, credit or prompt figure for any tier, and its own API returns a percentage with the total zeroed out. But the two halves can be put together: the console reports tokens consumed, the API reports the percentage left. Divide one by the other and the denominator falls out.

Tier

Ultra ($120/month)

Weekly quota

7.4B tokens

Monthly

32B tokens

Effective

$0.0038 /M

Method. The console reports tokens consumed; the API reports the percentage of the weekly window still left. Dividing one by the other gives the denominator the vendor never publishes. Measured 2026-07-18 against the plan window 13-20 Jul: 7.68B tokens consumed over a rolling 7 days against 6% of the window remaining. Anyone can repeat this on their own account; it costs nothing and spends no quota.

Confidence. Derived from a SINGLE account; not an official figure. Robust to the API's integer-percentage resolution (worth about 0.05B) and to the rolling-versus-fixed window mismatch (7.2-7.7B). One caveat that cannot be resolved from outside: the internal weighting of cached input, fresh input and output has not been reverse-engineered, so this figure is valid at the observed ~92% cache-hit rate and should not be extrapolated to a very different one. Note also that the accounting has already changed once: cached tokens formerly did not count against the quota and now do, which raised the effective cost sharply and was not announced.

For scale: at list price that traffic would cost $0.525 per million blended, and about $0.06 per million once caching is working. The subscription lands near $0.004 — roughly 16x cheaper than pay-per-token even after caching, and two orders of magnitude under list.

Why this had to be measured. Cached tokens did not always count against this quota. At some point they began to, the effective cost rose sharply, and nothing was announced — the subscription simply started buying less. The internal weighting of cached input against fresh input and output remains un-reverse-engineered, so the figure above holds at the observed ~92% cache-hit rate and should not be stretched far beyond it. A number a vendor does not publish is also a number it can change quietly.

07 ProvidersBy provider Gateways & hosts Price comparison

By provider

One page per lab: its models, what they cost, the endpoints serving them and at what precision, plus its own plan if it sells one.

Agnes AIrequests

provider

Token Plan (Starter / Plus / Pro) · from $4/mo

AkashML

serves 5

No subscription of its own

Alibaba / Qwenrequests

16 models · serves 18

Coding Plan (Pro) · from $50/mo

Amazon Bedrock

serves 26

No subscription of its own

Ambient

serves 3

No subscription of its own

Anthropicopaque

6 models · serves 9

Claude Pro / Max 5x / Max 20x · from $20/mo

Arcee AI

serves 1

No subscription of its own

Arcee Ai

1 model

No subscription of its own

AtlasCloudrequests

serves 20

Coding Plan (Starter / Lite / Plus / Max / Ultra / Enterprise) · from $10/mo

Augment Codedollars

provider

Augment (Business / Enterprise) — Auggie CLI + Cosmos · from $100/mo

Azure

serves 42

No subscription of its own

Baidu

serves 6

No subscription of its own

BaseTen

serves 9

No subscription of its own

Cerebras

serves 3

No subscription of its own

Chutesopaque

serves 6

Plus / Pro · from $10/mo

Claude Platform on AWS

serves 1

No subscription of its own

Cloudflare

serves 8

No subscription of its own

Cohere

2 models · serves 2

No subscription of its own

CoreWeave

serves 13

No subscription of its own

Crusoe

serves 6

No subscription of its own

Cursordollars

provider

Pro / Pro+ / Ultra · from $20/mo

Decart

serves 3

No subscription of its own

DeepInfra

serves 49

No subscription of its own

DeepSeek

7 models · serves 2

No subscription of its own

DigitalOcean

serves 12

No subscription of its own

Fireworks

serves 10

No subscription of its own

Friendli

serves 4

No subscription of its own

GMICloud

serves 11

No subscription of its own

GitHubdollars

provider

Copilot (Pro / Pro+ / Max) · from $0/mo

Googleopaque

12 models · serves 48

AI Plus / Pro / Ultra · from $5/mo

Google AI Studioopaque

serves 18

AI Plus / Pro / Ultra · from $5/mo

Groq

serves 5

No subscription of its own

IBM

1 model

No subscription of its own

Inception

1 model · serves 1

No subscription of its own

Inceptron

serves 4

No subscription of its own

InclusionAI

3 models

No subscription of its own

Io Netrequests

serves 3

IO Intelligence (Standard / Professional / Developer) · from $0/mo

Kwaipilot

1 model

No subscription of its own

Mancer 2

serves 3

No subscription of its own

Mara

serves 2

No subscription of its own

Meituan

1 model

No subscription of its own

Meta

5 models · serves 2

No subscription of its own

MiniMaxopaque

2 models · serves 3

Token Plan (Plus / Max / Ultra) · from $20/mo

Mistralopaque

7 models · serves 4

Le Chat Free / Pro / Team (Mistral Vibe bundled) · from $0/mo

Modal

serves 1

No subscription of its own

ModelRun

serves 3

No subscription of its own

Moonshotopaque

4 models · serves 5

Moderato / Allegretto / Allegro / Vivace · from $19/mo

Moonshot / Kimiopaque

provider

Moderato / Allegretto / Allegro / Vivace · from $19/mo

Morphdollars

serves 7

Usage-based (Free / Pay-per-token / Scale flat-rate) · from $0/mo

Nebius

serves 5

No subscription of its own

Nex AGI

1 model · serves 1

No subscription of its own

NextBit

serves 2

No subscription of its own

Novita

serves 33

No subscription of its own

Nvidia

5 models · serves 1

No subscription of its own

Ollamaopaque

provider

Ollama Cloud (Free / Pro / Max) · from $0/mo

OpenAIopaque

21 models · serves 35

ChatGPT Plus / Pro 5x / Pro 20x · from $0/mo

OpenInference

serves 2

No subscription of its own

Parasail

serves 20

No subscription of its own

Phala

serves 15

No subscription of its own

Qoderdollars

provider

Qoder (Free / Pro / Pro+ / Ultra) · from $0/mo

Relace

serves 1

No subscription of its own

Sail Research

serves 4

No subscription of its own

SambaNova

serves 4

No subscription of its own

SiliconFlow

serves 21

No subscription of its own

StepFun

1 model · serves 1

No subscription of its own

StreamLake

serves 14

No subscription of its own

Syntheticdollars

provider

Synthetic (Standard) · from $30/mo

Tencent

1 model · serves 1

No subscription of its own

Thinking Machines

1 model

No subscription of its own

Together

serves 15

No subscription of its own

Upstage

2 models · serves 2

No subscription of its own

Venicedollars

serves 21

Venice (Free / Pro / Pro Plus / Max) · from $0/mo

Wafer

serves 5

No subscription of its own

Windsurf (Cognition/Devin)opaque

provider

Pro / Max / Teams · from $0/mo

Xiaomitokens

2 models · serves 2

MiMo Token Plan (Lite / Standard / Pro / Max) · from $6/mo

Z.ai / Zhipurequests

4 models · serves 4

GLM Coding (Lite / Pro / Max) · from $18/mo

opencodedollars

provider

opencode Go · from $10/mo

xAI

4 models · serves 16

No subscription of its own

Gateways, resellers & inference hosts

Not the lab, and not the ranking

The places that sell a model without being the lab that made it, kept out of the leaderboard and the first-party price columns on purpose — because the distinction they blur is the useful part. An inference host serves open-weight models on its own GPUs, so its price is a real price for that model. A router forwards to providers with little markup. A reseller proxies someone else’s closed API (GPT, Claude) and reprices it — its number is the vendor’s minus a cut, and its “X% cheaper” is a claim against a list price, only meaningful if you can verify the effective rate before paying. The reseller/router class is detailed below; the full roster of every inference host we track (StreamLake, Together, Groq, SambaNova & 39 more) is at the end of the section.

Service	Origin	Type	Serves	Billing unit	Converts to tokens?	Verifiable pre-pay?	Note
CometAPI	🇭🇰 Hong Kong SAR	reseller	GPT-5.6, Claude Sonnet 5 / Fable 5 / Opus 4.8, Gemini 3.6/3.5, DeepSeek V4 Pro, Kimi K3 (500+ models)	tokens (official models) + per-image/clip for media	partial	partial	The most transparent of the resellers here: the 0.8x formula is stated outright, so the effective token price is checkable against the vendor's own list. Still a proxy of closed APIs, so the price is theirs minus 20%, not a first-party price, and no monthly plan.
AIMLAPI	🇪🇪 Estonia	reseller	GPT-5.6 Sol, DeepSeek V4 Flash, Gemini 3.6 Flash, coder models (600+); Claude not listed	tokens (per 1M)	full	partial	Prices per million tokens directly, so no credit layer to decode — but it publishes no discount claim, and Anthropic models are absent, so it is not a full first-party substitute.
APIMart	🇭🇰 Hong Kong SAR	reseller	GPT-5, Claude Sonnet 4.5, DeepSeek V3.2, Gemini + heavy image/video (Sora, Midjourney, Flux)	credits	none	false	OpenAI-compatible (base api.apimart.ai/v1), but billed in credits with no published credit-to-token conversion, so the advertised saving cannot be verified before paying. This is the entry that prompted the section.
EURouter	🇳🇱 Netherlands	router	Claude, GPT, Mistral, DeepSeek, Qwen, Codestral (100+); EU data residency	tokens + published markup	full	yes	A Netherlands-based router selling on EU data residency and GDPR rather than price. Notable as the transparent opposite of the credit resellers: it publishes its exact markup per tier and bills real tokens, so the effective price is checkable up front. Base www.eurouter.ai/api/v1.
ShareAI	—	inference-host	open-weight models only (Llama 4, GPT-OSS, Qwen, 150+) on a decentralized GPU grid	tokens	full	partial	Not a closed-API reseller but a decentralized inference marketplace: idle GPUs serve open weights, so its prices are real prices for those models. No frontier closed models (no GPT/Claude), so limited for coding against the leaders. Operator HeyShare SRL; country not confirmed.
APIMaster	—	reseller	OpenAI, Claude, DeepSeek via "fingerprint-verified channels"; works with Cursor / Claude Code	tokens (shared balance across channels)	none	false	OpenAI-compatible, but the headline discounts are far beyond a normal reseller margin and no markup is published, only per-"channel" prices that vary. Discounts that large on a closed API usually mean grey-market or quota-farmed capacity, not a straight resale. Company location undisclosed. Treat the savings claim with caution.
TokenMix	🇭🇰 Hong Kong SAR	reseller	Claude Opus 4.7, GPT-5, DeepSeek V4, Qwen, Gemini, Llama 4 (multi-provider)	tokens (per-model)	partial	partial	Token Limited (Hong Kong). Pay-per-token with per-model pricing and an OpenAI-compatible base (api.tokenmix.ai/v1). No markup formula published, but it bills tokens rather than opaque credits. $1 minimum top-up.
CrofAI	—	inference-host	open-weight coding models (Kimi, GLM, DeepSeek, Qwen, MiniMax, Gemma) at lowest rate	tokens (pay-per-token)	full	yes	Cheap OpenAI-compatible inference (crof.ai/v1) for open weights only, no markup, pure pay-per-token -- its own page now states \u201cno subscriptions, no minimums, only the tokens you actually burn\u201d (the tiered Hobby/Max plans some third-party listings still show are gone). One of the resellers this repo excluded from first-party price resolution earlier (as \u201ccrof\u201d); it belongs here, not in the leaderboard prices. Country not confirmed. checked 2026-07-22.
nano-gpt	—	router	GPT, Claude, DeepSeek, Qwen + open models (1000+), OpenAI-compatible	tokens (list price)	full	yes	A no-markup, privacy-first router: it passes provider list prices straight through (GPT-5.5 at $5/$30, same as OpenAI) with no deposit fee, and takes crypto from $0.10 with no KYC. The honest end of the spectrum -- the price is the vendor\u2019s, verifiable up front -- sold on privacy rather than a discount. Country not confirmed.
Fal.ai · Replicate · Wavespeed AI · PiAPI · EachLabs	—	reseller	generative image / video / audio (Midjourney, Flux, Kling, etc.)	per-second / per-run / credits	n/a	n/a	Named by APIMart as its competitors. All are media-generation gateways, not coding-LLM providers, so they are recorded here only to say they were checked and are out of scope — not to rank.
Featherless AI	🇸🇬 Singapore	inference-host	open-weight models only (40k+ pulled from Hugging Face) — DeepSeek V4 Pro/Flash, GLM-5.2, Qwen3, Kimi K2.x, Llama	flat monthly subscription by concurrency (not tokens)	none	partial	Serverless host of open weights (Singapore-founded by the RWKV team, AMD/Airbus-backed). Unusual model: you buy concurrency, not tokens — Premium/Chat $25/mo (4 units, 32K context), Agent $100/$200 (8 units, 256K). Overflow returns HTTP 429, no overage. It does serve DeepSeek V4 and GLM-5.2, but there is no per-token price to compare and no quantization disclosed, so cost per unit of real work is not derivable up front. A Feather Per-Request credit plan exists but its rate card is unpublished.
Nscale	🇬🇧 United Kingdom	inference-host	open-weight serverless — gpt-oss, Qwen2.5-Coder / Qwen3, Llama 4, Devstral, DeepSeek R1 distills (no full DeepSeek V4, no GLM, no Kimi)	tokens (per 1M, pay-as-you-go)	full	partial	UK-domiciled European "sovereign" AI cloud (London HQ; data centres in Norway / UK / Iceland). Two products: a GPU-rental hyperscaler (per-hour, its main business) and a smaller per-token serverless API. The serverless roster is mid-tier for coding — Qwen2.5-Coder, Devstral and gpt-oss — but the current open flagships (DeepSeek V4, GLM-5.2, Kimi) are absent, and DeepSeek is offered only as R1 distills. Quantization not disclosed; the full catalogue sits behind a signup gate.
OVHcloud AI Endpoints	🇫🇷 France	inference-host	open-weight, EU-sovereign — Llama, Mistral (incl. Codestral / Small 3.2), Qwen3 Coder, gpt-oss, DeepSeek R1 distill (no DeepSeek V4, no GLM, no Kimi)	tokens (per 1M, priced in EUR)	full	yes	France-based (OVHcloud, Roubaix), GDPR / EU data residency, served from Gravelines; states customer data is never used to train models. Open per-token catalogue in EUR (from ~€0.04/M). Relevant coding options are Qwen3 Coder 30B and gpt-oss-120b — the current open coding SOTA (DeepSeek V4, GLM-5.2, Kimi) is not onboarded. Prices are EUR, so the USD-normalised cost floats with FX. Quantization not disclosed.
Scaleway Generative APIs	🇫🇷 France	inference-host	open-weight, EU-sovereign — GLM-5.2, Qwen3 Coder, Mistral (Devstral 2, Large 3, Small 3.2), Llama, gpt-oss, MiniMax-M2.5 (DeepSeek R1 distills only, no DeepSeek V4, no Kimi)	tokens (per 1M, priced in EUR)	full	yes	France-based (Scaleway, Iliad group), served from Paris only; GDPR / EU residency and states it does not read or reuse prompt/output content. The most relevant of the HF-partner hosts for coding: it serves GLM-5.2 and Qwen3 Coder and — unusually — DISCLOSES quantization openly via model-id suffixes (:fp8, :bf16, :fp4, :int4, :awq). Prices are EUR (normalise to USD with FX). DeepSeek is only R1 distills, not V4. Single region (Paris), so no failover.
Public AI	—	inference-host	public-good / sovereign open-weight models only — Apertus, SEA-LION, EuroLLM, Bielik, Olmo (no mainstream or coding models)	free (donated compute)	n/a	partial	A non-profit "public utility for AI" (Public AI Inference Utility), serving publicly-funded sovereign models on donated / partner clusters, also routed free via Hugging Face. Currently free with a ~20 req/min limit, but "free" is explicitly provisional and there is no price sheet or SLA. The roster is sovereign / multilingual general-instruct models (Apertus, SEA-LION, EuroLLM) with no DeepSeek / GLM / Qwen-base / Llama / gpt-oss and no dedicated coding model, so it is recorded for completeness, not as a coding option. Legal HQ not stated (Swiss / EU centre of gravity).
Hugging Face Inference Providers	🇺🇸 United States	router	open-weight models across partner hosts (Cerebras, Groq, Together, Fireworks, Novita, DeepInfra, Scaleway, Z.ai …) through one HF token	tokens (pass-through provider rate)	full	yes	A router, not a host: one HF token routes to leading inference providers, OpenAI-compatible (router.huggingface.co/v1), with no HF markup — you pay the partner's rate. The provider is chosen by policy suffix (:fastest default, :cheapest, :preferred, or an explicit :provider). A small monthly credit ($0.10 free / $2 on PRO at $9/mo) then pay-as-you-go. It serves whatever its partners serve (open weights only), so price and quantization come from the chosen partner, not HF. Closed models (GPT-5.x, Claude) are not routed.
NVIDIA Build	🇺🇸 United States	inference-host	100+ open-weight models via an OpenAI-compatible API on NVIDIA's own GPUs — Nemotron, Llama, Qwen, gpt-oss, DeepSeek and GLM (Zhipu) families	free developer credits (production via NIM / per-token)	full	partial	NVIDIA's hosted API catalogue (build.nvidia.com), OpenAI-compatible — a genuinely generous free developer tier over 100+ open models on NVIDIA's own hardware, no credit card. It is not a per-token storefront: the hosted catalogue is free for prototyping (rate-limited ~40 RPM), and production means self-hosting the NIM microservice under NVIDIA AI Enterprise (~$4,500/GPU/year, with a 90-day free eval) or a quoted per-token range of ~$0.10-$10/M. Serves recent open models — strong on its own Nemotron line — but not always the very latest frontier (newest DeepSeek/GLM). US company.
RunPod	—	inference-host	Public Endpoints — a LIMITED set of pre-deployed OpenAI-compatible models, image/video/audio-first (Flux, Qwen Image, WAN 2.5, Kling, SORA 2) with only a few text/code models (Qwen3 32B, IBM Granite, Moonshot Kimi). Any other open weight you self-deploy on Serverless (per-second GPU) or Pods (per-hour GPU).	tokens	full	partial	Primarily a GPU-rental / serverless platform, but it DOES expose an OpenAI-compatible inference API via Public Endpoints (api.runpod.ai/v2/...), so a coding agent can point at it. Caveats for coding: the managed text catalogue is thin — Qwen3 32B is the main code-capable one, at a steep $10 per 1M tokens — and the platform is image/video-first. It is NOT an OpenRouter provider (no routed named-model prices); for arbitrary weights you self-deploy and pay GPU-time (Serverless per-second, Pods per-hour), not per-named-token. Sits between raw GPU rental (Vast/Lambda) and a full per-token host (DeepInfra/Together).

The open-weight inference hosts and the router are already priced per model in the endpoint tables above (the “where to buy” view per provider). This table exists for the reseller class, which those tables do not capture because a reseller’s repriced number is not a first-party price. Media-only gateways are listed to record that they were checked and are out of a coding site’s scope. Data in gateways.yaml.

What common models cost through every host

Input / output per million tokens for five reference models, against the vendor’s own list price — the resellers at the top (a fixed set we researched), then every OpenRouter host that serves them, with its real price and the precision it runs at. Two findings. “Gateway” does not mean “cheaper”: AIMLAPI and EURouter charge more than official, only CometAPI/TokenMix undercut it, and the credit/channel resellers can’t be placed on the scale at all. And for the open models the price is only half the story — the other half is quantization: the closed models (GPT-5.6, Claude) are first-party only, so hosts show “—”, while a cheaper DeepSeek here may be fp4, not fp8. Verified 2026-08-14.

Service	DeepSeek V4 Pro · fp8 native	DeepSeek V4 Flash · fp8 native	GLM-5.2 · fp8 native	GPT-5.6 Sol	Claude Sonnet 5	Basis
Official (first-party)	$0.435 / $0.87	$0.14 / $0.28	$1.4 / $4.4	$5 / $30	$2 / $10	the list price
CometAPI	$0.35 / $0.70	$0.11 / $0.22	$1.12 / $3.52	$4 / $24	$1.60 / $8	0.8x official; quant undisclosed
AIMLAPI	not listed	$0.18 / $0.36	—	$6.50 / $39	not served	per-model list; dearer on GPT, no Claude
APIMart	opaque	opaque	opaque	opaque	opaque	credits, no conversion
EURouter	+3-15%	+3-15%	+3-15%	+3-15%	+3-15%	markup by tier on official
ShareAI	open, per-token	open, per-token	open, per-token	not served	not served	open weights; quant undisclosed
APIMaster	opaque	opaque	opaque	opaque	opaque	"channels", up to 90% off claimed
TokenMix	$0.41 / $0.83	$0.13 / $0.26	$2.24 / $7.82	$4.75 / $28.50	Fable 5 only	per-model; GLM (Fast Preview) dearer, quant undisclosed
CrofAI	$0.35 / $0.80 · Q8_0	$0.12 / $0.21 · Q4_0	$0.30 / $1.05 · Q8_0	not served	not served	open weights, per-token; publishes quant (Q8_0 ≈ 8-bit, Q4_0 ≈ 4-bit)
nano-gpt	$0.435 / $0.87	$0.14 / $0.28	$1.4 / $4.4	$5 / $30	$2 / $10	list price, no markup; routes so quant varies
Featherless AI	served · flat sub, no per-token	served · flat sub, no per-token	served · flat sub, no per-token	not served	not served	$25-200/mo by concurrency, unlimited tokens; quant undisclosed
Nscale	not served	not served (R1 distills only)	not served	not served	not served	open-weight per-token; Qwen2.5-Coder 32B $0.06/$0.20, gpt-oss-120b $0.10/$0.40, Devstral Small $0.10/$0.30
OVHcloud AI Endpoints	not served	not served (R1 distill only)	not served	not served	not served	open-weight per-token EUR; Qwen3 Coder 30B ~€0.07/€0.26, gpt-oss-120b €0.08/€0.40, Mistral Small 3.2 €0.09/€0.28
Scaleway Generative APIs	not served	not served (R1 distills only)	€1.80 / €5.50 · fp8	not served	not served	open-weight per-token EUR, quant disclosed; GLM-5.2 €1.80/€5.50, Qwen3 Coder 30B €0.20/€0.80, Mistral Small 3.2 €0.15/€0.35
Public AI	not served	not served	not served	not served	not served	free sovereign open models; no coding model, quant undisclosed
Hugging Face Inference Providers	partner rate (pass-through, if a partner serves it)	partner rate (pass-through, if a partner serves it)	partner rate (pass-through, e.g. via Scaleway / Z.ai)	not served (open weights only)	not served (open weights only)	pass-through partner price, no markup; +$0.10/$2 monthly credit
NVIDIA Build	not served (older DeepSeek gen)	not served	not served (GLM-4 family)	not served	not served	free dev tier (credits, ~40 RPM); production via NIM / AI Enterprise
RunPod	not served (self-deploy on Serverless/Pods)	not served (self-deploy)	not served (self-deploy)	not served (closed model)	not served (closed model)	Public Endpoint text — Qwen3 32B $10.00/1M tokens; image Flux Dev $0.02/megapixel; video WAN 2.5 $0.50/5s. Self-deploy = GPU-time (Pods per-hour, Serverless per-second).
AkashML	—	$0.14 / $0.28^fp8	$0.77 / $2.42^fp8	—	—	OpenRouter · live
Alibaba / Qwen	—	—	$0.97 / $3.04^fp8	—	—	OpenRouter · live
Amazon Bedrock	—	—	—	$5.50 / $33.00^?	$2.00 / $10.00^?	OpenRouter · live
Ambient	—	$0.14 / $0.28^fp4	$1.05 / $4.40^fp8	—	—	OpenRouter · live
Anthropic	—	—	—	—	$2.00 / $10.00^?	OpenRouter · live
AtlasCloud	—	$0.14 / $0.28^fp4	$1.26 / $3.96^fp8	—	—	OpenRouter · live
Azure	—	—	—	$5.00 / $30.00^?	$2.00 / $10.00^?	OpenRouter · live
Baidu	—	$0.08 / $0.16^fp8	$0.49 / $1.54^fp8	—	—	OpenRouter · live
BaseTen	$1.32 / $3.96^fp4	$0.13 / $0.26^fp8	$1.40 / $4.40^fp8	—	—	OpenRouter · live
Cloudflare	$1.32 / $3.96^?	$0.14 / $0.28^?	$1.40 / $4.40^?	—	—	OpenRouter · live
CoreWeave	—	$0.13 / $0.28^fp8	$0.76 / $2.42^fp4	—	—	OpenRouter · live
Crusoe	—	—	$1.40 / $4.40^fp8	—	—	OpenRouter · live
Decart	—	$0.08 / $0.16^fp4	$0.77 / $2.56^fp4	—	—	OpenRouter · live
DeepInfra	—	$0.08 / $0.18^fp4	$0.75 / $2.40^fp4	—	—	OpenRouter · live
DeepSeek	$0.43 / $0.87^?	$0.14 / $0.28^fp8	—	—	—	OpenRouter · live
DigitalOcean	—	$0.08 / $0.25^?	$0.63 / $1.98^?	—	—	OpenRouter · live
Fireworks	$1.32 / $3.96^?	$0.14 / $0.28^?	$1.40 / $4.40^?	—	—	OpenRouter · live
Friendli	—	—	$1.40 / $4.40^?	—	—	OpenRouter · live
GMICloud	$1.22 / $2.44^fp8	$0.08 / $0.17^fp8	$0.74 / $2.33^fp8	—	—	OpenRouter · live
Google	—	—	—	—	$2.00 / $10.00^?	OpenRouter · live
Inceptron	—	$0.13 / $0.28^fp4	$0.75 / $2.90^fp4	—	—	OpenRouter · live
Io Net	—	$0.15 / $0.32^fp8	—	—	—	OpenRouter · live
Mancer 2	—	$0.17 / $0.50^fp8	—	—	—	OpenRouter · live
Morph	—	$0.14 / $0.28^bf16	$1.10 / $4.10^fp4	—	—	OpenRouter · live
Novita	$1.32 / $3.96^fp8	$0.14 / $0.28^fp8	$0.74 / $2.33^fp8	—	—	OpenRouter · live
OpenAI	—	—	—	$2.50 / $15.00^?	—	OpenRouter · live
OpenInference	—	$0.08 / $0.18^fp4	—	—	—	OpenRouter · live
Parasail	—	$0.14 / $0.28^fp8	$1.40 / $4.40^fp4	—	—	OpenRouter · live
Phala	—	$0.20 / $0.40^?	$1.13 / $3.00^fp8	—	—	OpenRouter · live
Relace	—	$0.14 / $0.28^fp4	—	—	—	OpenRouter · live
Sail Research	—	$0.09 / $0.18^fp4	$0.50 / $3.15^fp8	—	—	OpenRouter · live
SiliconFlow	$1.32 / $3.96^fp8	$0.14 / $0.28^fp8	$1.19 / $3.74^fp8	—	—	OpenRouter · live
StreamLake	—	$0.09 / $0.18^fp8	$0.75 / $2.37^fp8	—	—	OpenRouter · live
Together	—	$0.14 / $0.28^?	$1.40 / $4.40^?	—	—	OpenRouter · live
Venice	—	$0.17 / $0.35^?	$1.40 / $4.40^fp8	—	—	OpenRouter · live
Wafer	—	$0.28 / $0.56^?	$1.26 / $3.96^fp4	—	—	OpenRouter · live
Z.ai / Zhipu	—	—	$1.40 / $4.40^fp8	—	—	OpenRouter · live

Inference hosts & serving providers (44)

Every provider that serves a tracked model’s endpoint — open-weight hosts, cloud platforms, and labs serving their own models. These are priced per model in the “where to buy” tables on each provider page (from OpenRouter’s per-endpoint data, with quantization), which is why they are a roster here rather than a price grid: a host’s price depends on the specific model and precision, not a single rate. Click through for the per-model breakdown.

Provider	Origin	Site
AkashML	🇺🇸 United States	akashml.com
Alibaba	🇨🇳 China	www.alibabacloud.com/product/machine-learning
Amazon Bedrock	🇺🇸 United States	aws.amazon.com/bedrock/
Ambient	—	www.ambient.xyz
Anthropic	🇺🇸 United States	www.anthropic.com
AtlasCloud	—	www.atlascloud.ai
Azure	🇺🇸 United States	azure.microsoft.com/products/ai-foundry
Baidu	🇨🇳 China	cloud.baidu.com
BaseTen	🇺🇸 United States	www.baseten.co
Cerebras	🇺🇸 United States	www.cerebras.ai
Chutes	—	chutes.ai
Cloudflare	🇺🇸 United States	developers.cloudflare.com/workers-ai/
Crusoe	🇺🇸 United States	crusoe.ai
Decart	🇮🇱 Israel	www.decart.ai
DeepInfra	🇺🇸 United States	deepinfra.com
DeepSeek	🇨🇳 China	www.deepseek.com
DigitalOcean	🇺🇸 United States	www.digitalocean.com/products/gradient
Fireworks	🇺🇸 United States	fireworks.ai
Friendli	🇰🇷 South Korea	friendli.ai
GMICloud	🇺🇸 United States	www.gmicloud.ai
Google	🇺🇸 United States	cloud.google.com/vertex-ai
Groq	🇺🇸 United States	groq.com
Inceptron	—	www.inceptron.io
Io Net	🇺🇸 United States	io.net
Ionstream	—	ionstream.ai
Minimax	🇨🇳 China	www.minimax.io
Mistral	🇫🇷 France	mistral.ai
Moonshot AI	—	www.moonshot.ai
Morph	🇺🇸 United States	morphllm.com
Nebius	🇳🇱 Netherlands	nebius.com
Novita	—	novita.ai
Parasail	🇺🇸 United States	www.parasail.io
Phala	—	phala.network
RunPod	—	www.runpod.io
SambaNova	🇺🇸 United States	sambanova.ai
SiliconFlow	🇨🇳 China	www.siliconflow.com
StreamLake	🇨🇳 China	www.streamlake.ai
Together	🇺🇸 United States	www.together.ai
Venice	🇺🇸 United States	venice.ai
Wafer	—	wafer.ai
WandB	🇺🇸 United States	wandb.ai
xAI	🇺🇸 United States	x.ai
Xiaomi	🇨🇳 China	platform.xiaomimimo.com
Z.AI	🇨🇳 China	z.ai

08 Run localPick of the week By VRAM tier

Pick of the week

What I actually run

2026-08-14

What I actually run at home

Qwen3.8-27B is here, but at Q3 on 16 GB it is not my upgrade yet · staying on the old Qwen until I get a bigger card

Last week I said we were not yet at productive local coding on the hardware most of us have, and this week put it to the test. Qwen3.8-27B dropped, a dense 27B under Apache-2.0, and the exciting part is speed: with MTP speculative decoding through llama-server it runs at 36 tok/s on my 16 GB card, twice what it does without spec-decode, and finally interactive. But here is the honest catch. To fit 16 GB you run it at Q3 (about 14 GB), and at Q3 it scores 27% on our BCB-Hard, which is actually below what I already run locally: the ternary Qwen3.6-27B rebuild is at 35%, and Gemma ranks higher still. The full model at good precision is 37 to 43%, so the 3-bit quant that fits my card costs about 10 points, and that gap is the whole story. So I am not switching yet. I will keep running my old Qwen until I have a card that fits Q4 or Q6, where the 27B actually pulls ahead. That is a spend not everyone can make. If you can, the dense 27B at Q4 or Q6 is what to put on the new card. If you cannot, the real hope is not this dense model but the small A3B MoE Qwen will ship next, which runs fast on the 16 GB you already own. Until one of those lands, I am staying on my old Qwen.

What to run locally

Your GPU, not a bill

VRAM figures are real GGUF bytes and coding numbers are our own first-party runs where the AA index has none. The catalogue keeps growing; treat it as a solid guide, not gospel, and corrections are welcome.

The best coding models that fit each VRAM budget — cumulative, because a bigger card also runs everything smaller. Sizes are real GGUF bytes at ~4-bit (measured where a build exists), plus the KV cache the context needs — weights alone are not the VRAM figure the review videos quote. Ranked on one unified scale so the pick is realistic at every budget: our measured BigCodeBench-Hard (shown as % BCB) for the models we actually ran on this GPU, and the Artificial-Analysis coding index (shown as idx) for the big ones we can’t run locally — mapped onto the same axis by our measured fit (r = 0.86). So a multi-GPU tier correctly points at a GLM-5.2 or Kimi, while at 16 GB our own tests put gpt-oss-20b and the ternary Bonsai 27B on top — the very coders the index scores low or not at all. Reasoning models that run out of token budget locally are ranked on that measured floor, not their optimistic index. Quality still isn’t proportional to size — a well-built 12B beats a 20B. Each model links to how to run it.

≤8 GB

#	Model	Coder	Params	Weights ~Q4	Min VRAM @8K	Type
1	Gemma 4 12B	36% BCB	12B	7 GB	8 GB	dense
2	Ternary Bonsai 27B	36% BCB	27B	7 GB	8 GB	dense
3	Granite 4.1 8B	17% BCB	8B	5 GB	7 GB	dense

≤16 GB

Best pick unchanged from ≤8 GB: Gemma 4 12B still leads and needs only 8 GB. More VRAM here buys headroom and context, not a better coder — the larger open models score lower.

#	Model	Coder	Params	Weights ~Q4	Min VRAM @8K	Type
1	Gemma 4 12B	36% BCB	12B	7 GB	8 GB	dense
2	Ternary Bonsai 27B	36% BCB	27B	7 GB	8 GB	dense
3	Qwen3 14B	31% BCB	14B	9 GB	11 GB	dense

≤32 GB

#	Model	Coder	Params	Weights ~Q4	Min VRAM @8K	Type
1	Qwen3.6 27B	idx 53.7	27B	17 GB	19 GB	dense
2	Gemma 4 26B A4B	idx 39.3	26B	17 GB	19 GB	dense
3	Nemotron Cascade 2 30B A3B	idx 25.3	32B	25 GB	28 GB	MoE

≤80 GB

Best pick unchanged from ≤32 GB: Qwen3.6 27B still leads and needs only 19 GB. More VRAM here buys headroom and context, not a better coder — the larger open models score lower.

#	Model	Coder	Params	Weights ~Q4	Min VRAM @8K	Type
1	Qwen3.6 27B	idx 53.7	27B	17 GB	19 GB	dense
2	Gemma 4 26B A4B	idx 39.3	26B	17 GB	19 GB	dense
3	gpt-oss-120b	idx 30.4	117B	63 GB	70 GB	MoE

Data centre

#	Model	Coder	Params	Weights ~Q4	Min VRAM @8K	Type
1	DeepSeek V4 Flash	idx 69.1	291B	137 GB	151 GB	MoE
2	GLM-5.2	idx 68.8	753B	466 GB	544 GB	MoE
3	DeepSeek V4 Pro	idx 68.8	862B	850 GB	936 GB	MoE

Every model, ranked

All 53 models in the catalogue by coding index. Filter the Fits or Type column to narrow to your hardware, or sort any column. HumanEval* and BCB* (BigCodeBench-Hard) are first-party pass@1 scores we ran ourselves — HumanEval on 12 models, the harder BCB-Hard on 23, on the real GPU. Note how HumanEval saturates near the top while BCB-Hard spreads the field out — that gap is why we run both. A dimmed ~NN% est in the BCB column is an estimate from the coding index for a model we haven’t run — only shown where the index is inside the range we fitted (r = 0.86), never extrapolated wildly. t/s is the measured local decode speed on an RTX 4060 Ti. Click a measured % for how it was run.

Model	Coding idx	HumanEval*	BCB*	t/s	Params	Weights ~Q4	Min VRAM	Fits	Type
DeepSeek V4 Flash	69.1	—	—	—	291B	137 GB	151 GB	Data centre	MoE
GLM-5.2	68.8	—	—	—	753B	466 GB	544 GB	Data centre	MoE
DeepSeek V4 Pro	68.8	—	—	—	862B	850 GB	936 GB	Data centre	MoE
Motif-3 Beta	63.5	—	—	—	315B	196 GB	221 GB	Data centre	MoE
Kimi K2.6	61.8	—	—	—	1027B	584 GB	642 GB	Data centre	dense
Kimi K2.7 Code	60.8	—	—	—	1027B	495 GB	545 GB	Data centre	dense
MiMo-V2.5-Pro	60.2	—	—	—	1023B	630 GB	696 GB	Data centre	MoE
Nex-N2-Pro	59.1	—	—	—	397B	242 GB	266 GB	Data centre	dense
Hy3	58.8	—	—	—	299B	182 GB	203 GB	Data centre	MoE
GLM-5.1	55.8	—	—	—	754B	465 GB	521 GB	Data centre	MoE
Qwen3.6 27B	53.7	—	—	—	27B	17 GB	19 GB	≤32 GB	dense
Inkling	52.9	—	—	—	952B	571 GB est	629 GB	Data centre	dense
Nemotron 3 Ultra 550B	49.3	—	—	—	550B	359 GB	395 GB	Data centre	MoE
Qwen3.5 397B A17B	48.2	—	—	—	397B	488 GB	537 GB	Data centre	dense
Mistral Medium 3.5	46.9	—	—	—	128B	75 GB	82 GB	Data centre	dense
Qwen3.5 122B A10B	45.7	—	—	—	125B	77 GB	84 GB	Data centre	dense
Solar Open2 250B	44.7	—	—	—	250B	136 GB	151 GB	Data centre	MoE
Gemma 4 26B A4B	39.3	—	—	—	26B	17 GB	19 GB	≤32 GB	dense
Nemotron 3 Super 120B	37.7	—	—	—	120B	83 GB	92 GB	Data centre	MoE
gpt-oss-120b	30.4	—	—	—	117B	63 GB	70 GB	≤80 GB	MoE
Nemotron Cascade 2 30B A3B	25.3	—	—	—	32B	25 GB	28 GB	≤32 GB	MoE
EXAONE 4.5 33B	23.6	—	—	—	33B	20 GB	22 GB	≤32 GB	dense
Qwen3 235B A22B 2507	22.1	—	—	—	235B	142 GB	158 GB	Data centre	MoE
Gemma 4 31B	43.4	—	43%	—	31B	18 GB	20 GB	≤32 GB	dense
Qwen3-Coder 30B A3B	—	98%	42%	40 t/s	31B	19 GB	21 GB	≤32 GB	MoE
Qwen3 Next 80B A3B	17.4	—	~37% est	—	80B	49 GB	54 GB	≤80 GB	MoE
Gemma 4 12B	31.0	63%	36%	31 t/s	12B	7 GB	8 GB	≤8 GB	dense
Ling-3.0-flash	50.6	—	36%	—	124B	155 GB	176 GB	Data centre	MoE
Ternary Bonsai 27B	—	90%	36%	30 t/s	27B	7 GB	8 GB	≤8 GB	dense
Nemotron 3.5 Lightning	26.8	—	35%	—	30B	18 GB	20 GB	≤32 GB	MoE
Ornith-1.0 35B	—	—	35%	—	35B	22 GB	24 GB	≤32 GB	dense
North Mini Code 1.0	36.5	—	35%	—	30B	18 GB est	21 GB	≤32 GB	MoE
Laguna XS.2	—	—	34%	—	33B	21 GB	24 GB	≤32 GB	MoE
Qwen3.6 35B A3B	41.9	—	33%	—	35B	22 GB	24 GB	≤32 GB	dense
Qwen3 32B	15.3	—	~33% est	—	33B	20 GB	24 GB	≤32 GB	dense
Qwen3 14B	13.8	63%	31%	25 t/s	14B	9 GB	11 GB	≤16 GB	dense
JetBrains Mellum 2 12B	—	—	28%	—	12B	8 GB	9 GB	≤16 GB	MoE
Qwen3.8-27B	—	—	27%	36 t/s	27B	14 GB	15 GB	≤16 GB	dense
gpt-oss-20b	20.7	97%	26%	61 t/s	22B	12 GB	14 GB	≤16 GB	MoE
Llama 3.3 70B	11.9	—	~25% est	—	70B	43 GB	47 GB	≤80 GB	dense
Mistral Small 3.2 24B	12.5	83%	25%	11 t/s	24B	14 GB	16 GB	≤16 GB	dense
Nemotron Nano 9B V2	—	—	22%	—	9B	7 GB	9 GB	≤16 GB	dense
Granite 4.1 30B	10.4	—	~22% est	—	29B	17 GB	21 GB	≤32 GB	dense
Granite 4.1 8B	9.5	67%	17%	46 t/s	8B	5 GB	7 GB	≤8 GB	dense
Nanbeige 4.2-3B	—	81%	17%	9 t/s	4B	3 GB	4 GB	≤8 GB	dense
Ministral 3 14B	14.4	63%	14%	30 t/s	14B	8 GB	9 GB	≤16 GB	dense
Ornith-1.0 9B	—	—	12%	—	9B	6 GB	6 GB	≤8 GB	dense
Qwen3.5 9B	28.7	40%	12%	44 t/s	9B	6 GB	6 GB	≤8 GB	dense
Qwen3.5 4B	22.6	33%	12%	71 t/s	5B	3 GB	3 GB	≤8 GB	dense
Bonsai 8B	—	63%	6%	156 t/s	8B	1 GB	2 GB	≤8 GB	dense
Instella MoE 16B A3B	—	—	—	—	16B	10 GB	13 GB	≤16 GB	MoE
KAT-Coder V2.5 35B	—	—	—	—	35B	21 GB	24 GB	≤32 GB	dense
Laguna S 2.1	—	—	—	—	118B	73 GB	82 GB	Data centre	MoE

Each tier is everything that fits that budget, so the same strong small model can top several tiers — that repetition is the point. “Min VRAM” is weights plus KV cache at an 8K context; longer context or a heavier quant (Q8/fp16) needs more. A dash under coding index is unmeasured, never zero. Data in local.yaml.

It’s free

beta$0 to start

Beta. New view — free tiers change constantly, so amounts are approximate and “free for now” is a real caveat. Verify before relying on any of it.

Every way we track to run a capable model for nothing — a real free allowance, not a trial that needs a card. Ranked roughly by how much you get. The catch column is the honest part: rate limits, one-off vs monthly, or “free for now”.

Free API credits & allowances

Hosted, OpenAI-compatible APIs with a genuine free tier over open (and some closed) models.

Provider	What you get free	The catch
Public AI	unlimited, for now	donated compute, ~20 req/min, sovereign models only, may end
OpenRouter	15 `:free` routes today — gemma-4-26b-a4b-it, gemma-4-31b-it, gpt-oss-20b, laguna-s-2.1, laguna-xs-2.1, lfm-2.5-2.6b, nemotron-3-nano-30b-a3b, nemotron-3-nano-omni-30b-a3b-reasoning, nemotron-3-super-120b-a12b, nemotron-3-ultra-550b-a55b, nemotron-3.5-content-safety, nemotron-3.5-lightning, nemotron-nano-12b-v2-vl, nemotron-nano-9b-v2, north-mini-code	~50 requests/day (1,000 once you have topped up $10+), 20/min; the set rotates without notice — snapshot as of 2026-08-14
opencode Zen	Rotating free set (Big Pickle, DeepSeek V4 Flash, MiMo-V2.5, Hy3, Laguna S 2.1, Nemotron 3 Ultra, Nemotron 3.5 Lightning)	free through the opencode CLI; the set rotates without notice — as of 2026-08-14
Scaleway Generative APIs	1,000,000 tokens	one-off for new accounts, then per-token EUR
Google AI Studio (Gemini API)	~1,000–1,500 requests/day on Gemini 3 Flash / 2.5 Flash	no card, but free-tier prompts may be used to improve Google's models; ~5–15 req/min; Pro models removed from the free tier (Apr 2026)
Groq	~500K tokens/day per model + ~14,400 requests/day	no card; 30 req/min and 6K tokens/min caps; resets midnight UTC
Mistral La Plateforme	“Experiment” free tier — ~1B tokens/month	eval/prototyping only, not production; phone verification and a data-sharing opt-in
Cohere	1,000 API calls/month (trial keys)	non-production; ~20 req/min; data may be used to improve models
NVIDIA Build	1,000–5,000 API credits	~40 req/min, dev use; production needs NIM / AI Enterprise
Cloudflare Workers AI	10,000 Neurons/day (Free and Paid plans alike)	no card; resets 00:00 UTC; on the Free plan you are blocked until reset once it is spent
Nscale	~$5 signup credit	one-off; then per-token
OVHcloud AI Endpoints	sign-up credit + a few always-free models	new Public Cloud projects
Hugging Face Inference Providers	$0.10/mo free · $2/mo on PRO	monthly; pass-through provider rates after
Cerebras	Free trial ($5 credits) over all models (gpt-oss-120b, GLM-4.7, Gemma 4 31B)	no card for the trial; ~8K-token context cap; the widely-cited “sunset 2026-08-17” applies only to PREVIEW models in the paid Developer tier, not the free tier
SambaNova Cloud	Persistent free tier + $5 signup credit	the $5 credit expires in 3 months; per-model daily caps (some ~20 req/day)
Nebius Token Factory	No-card free trial credit (~$1) over open models (Llama-3.3-70B, Qwen3-235B)	rebranded from Nebius AI Studio in 2026; the free trial credit is now small (~$1), and a card is needed for the $25 minimum top-up

No longer free (don’t bother): Together (signup credit retired), Fireworks (~$1, testing only), Chutes (free tier ended Mar 15 2026), and GitHub Models (retired 2026-07-30). Listed so you don’t go looking.

Free forever (nothing to burn)

Run it yourself. Every open-weight model in Run local is free once you have the GPU — a 12B fits an 8 GB card and codes well, with no rate limit and no rotation. The rate-limited-free routes above (OpenRouter :free, opencode Zen) are also $0 with no expiry — the trade there is the provider’s rate cap and a set that rotates without notice.

Free subscription tiers

Products with a permanent $0 plan (limited, but real).

Provider	Free tier
GitHub	Free
Io Net	Standard
Mistral	Free
Morph	Free
Ollama	Free
OpenAI	Free
Qoder	Free
Venice	Free
Windsurf (Cognition/Devin)	Free

Data from gateways.yaml and plans.yaml. “Free for now” means exactly that — promotional tiers get cut (Tencent Hy3’s free window already closed), so treat any allowance as this week’s, not forever.

10 MethodWhat changed Historical data Sources Method

What changed

Since the snapshot of 2026-08-13.

GPT-5.6 Sol — provider count 6 → 7
Claude Opus 5 — provider count 9 → 10
GPT-5.6 Terra — provider count 6 → 7
Kimi K3 — cost per session $2.0189 → $2.0958
Kimi K3 — cache ratio 92.1% → 92.0%
GPT-5.5 — provider count 6 → 7
Claude Opus 4.8 — provider count 8 → 10
Grok 4.5 — cost per session $0.8088 → $0.8340
Grok 4.5 — cache ratio 86.8% → 86.7%
Nemotron 3.5 Lightning — context 1,048,576 → 1,000,000
Claude Sonnet 5 — provider count 8 → 9
GPT-5.6 Luna — provider count 6 → 7
GPT-5.4 — provider count 6 → 7
Gemini 3.6 Flash — input price $1.50 → $0.75
Gemini 3.6 Flash — output price $7.50 → $3.75
GLM-5.2 — input price $0.63 → $1.19
GLM-5.2 — output price $1.98 → $3.74
GLM-5.2 — cost per session $2.5186 → $2.5679
GLM-5.2 — cache ratio 80.9% → 80.6%
Qwen3.7 Max — cost per session $1.5437 → $1.5594
Claude Sonnet 4.6 — provider count 8 → 9
Kimi K2.6 — input price $0.58 → $0.56
Kimi K2.6 — output price $2.44 → $2.36
Kimi K2.6 — cost per session $0.7972 → $0.8180
Kimi K2.7 Code — input price $0.67 → $0.71
Kimi K2.7 Code — output price $3.40 → $3.50
Kimi K2.7 Code — cost per session $1.0773 → $1.0948
Kimi K2.7 Code — cache ratio 94.3% → 94.2%
Xiaomi MiMo-V2.5-Pro — cost per session $0.3798 → $0.3802
DeepSeek V4 Pro — input price $1.17 → $0.43
DeepSeek V4 Pro — output price $2.34 → $0.87
DeepSeek V4 Pro — coding index 59.4 → 68.8
DeepSeek V4 Pro — cost per session $0.4328 → $0.4384
DeepSeek V4 Pro — cache ratio 96.7% → 96.8%
DeepSeek V4 Pro — provider count 18 → 7
Tencent Hy3 — cost per session $0.3111 → $0.2951
MiniMax-M3 — cost per session $0.5314 → $0.5434
MiniMax-M3 — provider count 11 → 12
Xiaomi MiMo-V2.5 — cost per session $0.0698 → $0.0722
DeepSeek V4 Flash — input price $0.08 → $0.14

Historical data

The previous Model Speed Arena — weekly measured throughput runs against live subscriptions — is preserved at https://oldmsa.millaguie.net. Those runs are no longer updated: sustaining a weekly benchmark fleet was not viable. They remain the better source for measured tok/s under a real subscription, which no aggregator publishes.

Where the data comes from

Every figure on this page is somebody else’s measurement or publication. Nothing here is re-benchmarked locally, so each dataset is credited with what is taken from it and what that source can and cannot tell you.

Source	What this site takes from it	Nature
OpenRouter	List price, context window, and the per-provider quantization, uptime and endpoint list	public API
models.dev	Official provider catalog, tiered and cache pricing	public API
Artificial Analysis	Independent quality evaluations — coding index (which sets the ranking here), general intelligence index, TerminalBench-Hard and long-context reasoning — plus the only measured speed figures (tok/s, time to first token). The coding index is AA’s own weighted composite: we cannot recompute it, which is why a model AA has not tested cannot be ranked	independent lab
opencode	Observed production usage across its subscriber base: cost per session, tokens per session, cache ratio	aggregate telemetry
quota-sentinel	Which providers expose a readable quota endpoint	open source
Inference providers	Every serving provider links to its own site, each checked to resolve. Where a referral link exists it is used and marked, and the ranking is computed from price and precision alone regardless	neutral link
Vendor documentation	Subscription prices and quotas, reproduced as advertised. Each plan row links to its own source and carries the date it was checked	vendor-reported
Reseller gateways	Named and classified in the Gateways section, never priced into the leaderboard: a reseller reprices someone else’s API, so its number is not a first-party price	reseller
Third-party write-ups	Used only where a vendor publishes no figure at all. Kimi’s per-tier request counts are the case in point: its own page gives adjectives, so the numbers come from independent write-ups instead and the plan is marked low-confidence. Every such plan links the write-up it came from	unofficial
Logged-in console	Figures visible only inside an account, so a reader cannot click through to verify them: MiniMax’s remaining-quota percentage (from which the hidden quota here was derived), Google AI Plus’s euro price, and DeepSeek’s peak-pricing notice — which its public pricing page still does not mention	not publicly verifiable
Measured here	A small number of first-party runs, only for models no independent lab has scored: HumanEval, BigCodeBench-Hard and SWE-bench Verified. Own scaffold, so comparable between these runs and never to a public leaderboard. Code and redacted trajectories are in the repository	first-party

Two things worth carrying. Quality figures are tagged vendor or independent throughout, because the flattering numbers are almost always the vendor’s own. And observed-usage figures describe opencode’s subscribers, not the whole market — a model barely used there will have a noisy cost per session.

Method & limits

This site is an aggregator: price, context, per-provider quantization and uptime come from OpenRouter; official catalog pricing from models.dev; independent quality and speed from Artificial Analysis; observed production cost from opencode. Plan quotas were researched from vendors’ own documentation and are reproduced as advertised. With one exception, stated rather than hidden: a handful of models have no independent score anywhere, and for those a first-party run was done here (HumanEval, BigCodeBench-Hard, SWE-bench Verified) purely so the column is not blank. Those figures use our own scaffold, are labelled wherever they appear, and are never mixed into the leaderboard — which stays entirely Artificial Analysis’ coding index.

Two honest limits. Quota comparison is mostly impossible: only one provider publishes a quota that converts to tokens, and several publish nothing numeric at all. And vendor multipliers are not comparable — some are measured against a free tier, others against a paid one, some per session rather than per month. Plan pricing moves fast; every figure here is dated.

◆ Some links on this page are referral links: if you sign up through them this site may earn a commission, at no extra cost to you. Rankings, prices and measurements are taken from the sources listed under Method, and are not influenced by whether a provider has a referral programme.