Model Spend Arena2ND ED.
provider · 2026-08-14

Agnes AI

Official site · origin not confirmed

Plan

Token Plan (Starter / Plus / Pro)

Tier$/monthIncluded usage
Starter$4.00
Plus$10.00
Pro$50.00

console only partly convertible

As advertised. Starter 1,500 requests/5h and 15,000/week; Plus 7,500 and 75,000; Pro 30,000 and 300,000. All tiers: 4,000 images/day and 500s of video/day

What that actually means. Text quota is counted per request, images per generated image, video by generated seconds. Unusually precise for this market -- the numbers are published in the vendor's own public repository rather than buried in marketing -- but requests still do not convert to tokens, and no per-request cost is given. PRICES SIT BEHIND A LOGIN, which is why the public docs and GitHub catalog carry detailed quotas and no figures at all. Read from an authenticated account: $4, $10 and $50 a month, halved for the first month. With both halves in hand the plan becomes one of the few here where cost per unit of quota is computable. The weekly cap binds long before the 5-hour one -- it is roughly four times more restrictive -- so the real ceiling is 65,000 requests a month on Starter, 326,000 on Plus and 1.3M on Pro. That works out at $0.061, $0.031 and $0.038 per thousand requests: Plus is the sweet spot and Starter the worst value per request, which is the opposite of how the ladder is usually shaped. AND UNLIKE MOST HERE, THE REQUEST-TO-TOKEN CONVERSION IS OBTAINABLE. Every completion response carries a standard usage object with prompt, completion and total tokens, and those figures match the billing console exactly -- a call logged at 258 input and 3 output tokens in the console returned precisely that in its usage field. So while the vendor publishes no typical tokens-per-request figure, you can measure your own workload and build the conversion the quota table lacks. That is a materially different position from a plan where the relationship is withheld outright. OpenAI-compatible billing endpoints exist too: /v1/dashboard/billing/usage returns aggregate monetary consumption and /v1/dashboard/billing/subscription returns soft and hard limits. Neither returns remaining request quota, so the request counter itself stays console-only.

Where to check what is left. platform.agnes-ai.com console

Models covered. agnes-2.0-flash (text and vision, coding, tool calling), agnes-image-2.0/2.1-flash, agnes-video-v2.0

A unified gateway from Sapiens AI, OpenAI-compatible at apihub.agnes-ai.com/v1, rather than a lab or a coding IDE. It covers text, image and video on one balance, which is why its quota table has three separate units. Rate limits are tiered by account type rather than by plan: 30 requests per minute on a free key against 1,000 for a Token Plan key, and the docs distinguish "public request RPM" from "actual executable RPM", which is a candour most vendors skip. A free tier exists and reporting around the Agnes-2.5-Flash launch describes uncapped free access for agentic coding. That claim is press-sourced, and calling the API directly could not confirm it: the model is not reachable on a live key. The plan pages also quote throughput -- around 100 tokens per second, rising to 150 off-peak. Note the direction: where Qoder discounts the price off-peak and Z.ai surcharges the quota at peak, Agnes leaves the price alone and hands out extra speed. Same spare capacity, a third way of passing it on.

Measured here. No independent lab publishes a quality score for this vendor’s models, so they were run in this repo. HumanEval pass@1 95%, graded by executing the canonical unit tests. 19 of 20 on a stratified HumanEval subset, graded by executing the canonical unit tests. The single miss was not a wrong answer but a 120-second read timeout: of the 19 requests that returned, all 19 passed. Counting the timeout as a failure is the standard convention and gives 95%; on answers actually returned it was 100%. Latency was the free tier's real constraint -- median 44s per problem, range 9 to 120 -- not correctness. This is, as far as I can find, the only quality figure that exists for this model anywhere.

BigCodeBench-Hard. pass@1 35.0%. 14 of 40 on the same gold-gated BigCodeBench-Hard subset used elsewhere in this repo, graded by the official unittest suite. Three of the forty hit our own 300-second read timeout rather than answering; retried, all three returned and one passed, so the figure is 35.0% and not the 32.5% a naive scoring of those timeouts would give. That sits a few points below the full MiniMax-M3 (40.0%) measured the same way. Unlike HumanEval, this benchmark does not saturate, so it actually separates the models.

SWE-bench Verified. resolved 13/25 (52%), and every patch it submitted passed (16/16). 13 of 25 gold-gated SWE-bench Verified instances resolved (52%) under mini-swe-agent with an 80-step budget, each patch scored by the official harness. But the 80-step budget is the real limiter, not capability: Agnes failed to converge on 9 of the 25 (it kept exploring past the budget), yet 81% of the patches it DID submit passed. Re-running 4 of the non-converged instances with a 200-step budget resolved 2 more -- so its true ceiling here is materially higher than 52%. For comparison, Tencent Hy3 (a higher-quality model) run identically resolved 68% but converged far faster. This uses our own scaffold and step budget: comparable to our other runs, NOT to the public SWE-bench leaderboard.

checked 2026-08-13 · high (quotas, from the vendor's official repo; prices read from an authenticated account, since they are not shown publicly) · source

◆ 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.