Synthetic OpenAI-compatible

Coverage paused — Synthetic subscription not renewed for 2026-W22; historical data retained.

⚠️Heads up: Not benched in 2026-W22. Our Synthetic subscription lapsed and we chose not to renew. The provider stays listed because the page preserves historical context, but no fresh numbers will appear until the bench resumes. Want Synthetic covered? Sponsor a key — see About / sources or email millaguie [at] gmail [dot] com.

Synthetic (Synthetic Lab) runs open-source AI models in private, secure datacenters. They never train on user data or store API prompts and completions. The catalogue spans DeepSeek, GLM, Kimi, Llama, MiniMax, Qwen and more — any model vLLM supports. A flat-rate subscription at $30/mo covers all always-on models; usage-based PAYG is also available. Status (2026-W22): we did not renew the subscription this cycle, so no fresh numbers were recorded this week. Past results remain on this page for reference. If you'd like Synthetic kept on the weekly bench, lend us an API key with modest quota and we'll wire it back in — see the sponsorship offer below.

Strengths

Private datacenters — no data stored, no training on prompts
Flat $30/mo includes all always-on models
Broad open-source catalogue (DeepSeek, GLM, Kimi, Llama, MiniMax, Qwen)

When to use it

Coding agents that need privacy guarantees
Running many open-source models under one subscription
Switching between models without per-token billing surprises

Subscription plans

Plan	Price	Quota	Available
Subscription	$30/mo	500 messages / 5h · all models included · 1 concurrent req/model	yes
Usage-based	$0/mo	Pay-per-token · all models	yes

Notes: Subscription is $1/day ($30/mo). Each pack adds 1 concurrent request per model — buy more packs to scale. All always-on models are included in the subscription; no per-token charges. Usage-based PAYG is also available for enterprise. <strong>Active on MSA?</strong> Not this week — see availability warning.

Referral: Synthetic runs a referral program: sign up via the link above and both you and the referrer earn bonus API credits. We're also happy to receive an API key directly — see the availability warning above for the sponsorship path.

Models tested on Synthetic

Speed numbers below are specific to Synthetic's routing and hardware. The same model may appear on other providers' pages with different throughput.

Best tok/s observed on Synthetic per weekly snapshot (2 points).

Model	Best tok/s	Avg tok/s	Runs	Success	Longest output (chars)
hf:moonshotai/Kimi-K2.6	224.2	172.2	2	100%	4,093
hf:Qwen/Qwen3.5-397B-A17B	190.7	138.4	3	100%	3,393
hf:zai-org/GLM-4.7	173.7	164.1	3	100%	2,666
hf:zai-org/GLM-4.7-Flash	168.4	153.9	3	100%	2,828
hf:nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4	142.0	134.2	3	100%	3,959
hf:deepseek-ai/DeepSeek-R1-0528	135.2	132.9	3	100%	5,131
hf:deepseek-ai/DeepSeek-R1	131.2	122.4	3	100%	4,946
hf:zai-org/GLM-5.1	130.0	106.0	3	100%	2,053
hf:MiniMaxAI/MiniMax-M2.5	129.5	118.3	3	100%	3,304
hf:openai/gpt-oss-120b	127.3	93.9	3	100%	4,966
hf:moonshotai/Kimi-K2.5	126.2	126.2	3	33%	3,239
hf:zai-org/GLM-5	109.6	108.8	3	100%	3,894
hf:meta-llama/Llama-3.3-70B-Instruct	99.7	78.9	3	100%	2,549
hf:nvidia/Kimi-K2.5-NVFP4	98.1	98.1	3	33%	2,912
hf:Qwen/Qwen3-Coder-480B-A35B-Instruct	90.8	85.8	3	100%	2,613
hf:deepseek-ai/DeepSeek-V3	88.0	52.7	3	100%	5,068
hf:deepseek-ai/DeepSeek-V3.2	80.6	75.3	3	100%	3,549