Full 24-provider audit, every model re-checked against live APIs and official docs by 24 parallel researchers. Catalog goes 251 to 256 live free models. This audit also cleaned up several mistakes the 2026-09-15 audit introduced (dead or paid-only models that were wrongly re-added).
Removed (22)
- NVIDIA NIM (-5): deepseek-v4-flash-0731 died today (NVIDIA banner: deprecated 09/19, unsupported after 09/21), deepseek-v4-pro-0813 (410 Gone, EOL 09/14), qwen3-coder-480b-a35b-instruct (410 Gone since June - the 09-15 re-add resurrected a months-dead model), minimaxai/minimax-m3 (410 Gone, EOL 09/09) and minimaxai/minimax-m2.7 (410 Gone since 07/27).
- Groq (-5): the free tier collapsed. llama-3.1-8b-instant and llama-3.3-70b-versatile were shut down on 08/16 (the 09-15 re-add reverted a correct removal), minimax-m2.7 is enterprise-only on Groq (model_not_found on developer keys), and groq/compound + groq/compound-mini hit their shutdown date today. Groq is down to 3 free models: gpt-oss-120b, gpt-oss-20b, qwen3.8-27b.
- Cloudflare (-7): glm-5.3, glm-5.3-flash, glm-5.2, deepseek-v4-pro-0813, deepseek-v4-flash-0731, kimi-k2.6 and kimi-k2.7-code all now carry the "Paid access required: not available through standard Workers Free billing" badge on the official docs pages. They are unusable on the free 10k neurons/day tier, so they are out (this re-applies the 09-05 paid-only policy the 09-15 audit regressed).
- Z.ai (-2): glm-5.1 and glm-5.2 now silently redirect to GLM-5.3 on the Coding Plan, so the ids no longer serve a distinct free model.
- Google AI (-1): gemini-3.1-pro-preview re-removed - Pro models left the free tier around April 2026; the 09-15 re-add resurrected a paid-only model. The Gemini 2.5 family is still free but retires no earlier than 2026-10-16 (noted in-file).
- Cerebras (-1): qwen-3-235b-a22b-instruct-2507 was deprecated 2026-05-27; the official catalog now serves exactly 2 free models (gpt-oss-120b, qwen3.8-27b).
- SiliconFlow (-1): Qwen2.5-Coder-7B-Instruct was taken offline on 2026-03-17 (official release note); the 09-15 re-add resurrected a 6-months-dead id.
Added (27)
- Alibaba DashScope (+4): qwen3-coder-480b-a35b-instruct is back on the free billing page (1M-token free quota, page updated 09/20), plus new qwen3-coder-30b-a3b-instruct, qwen3-vl-plus and qwen3.8-omni-flash.
- Pollinations (+10): the flagship wave arrived - moonshotai/kimi-k3, deepseek-v4-pro, qwen3.8-max, gemini-3.1-pro-preview (free here even though paid on Google), openai/gpt-5.5, openai/gpt-6-astra, openai/gpt-5.6-luna, z-ai/glm-5.3-flash, nvidia/nemotron-3-ultra and qwen3-coder-next. Every id verified directly against the live /v1/models list (411 models).
- OVHcloud (+4): Qwen3-Coder-30B-A3B-Instruct, Mistral Small 3.2 24B, Mistral Nemo 12B and Mistral 7B v0.3 are back in the official AI Endpoints catalog.
- Google AI (+2): gemma-4-31b-it and gemma-4-26b-a4b-it joined the free tier.
- Kilo (+2): qwen3.8-27b:free and nemotron-3-nano-omni:free are new in the live free gateway list (20 free models now).
- OpenRouter (+1): qwen/qwen3.8-27b:free.
- Cloudflare (+1): @cf/aisingapore/gemma-sea-lion-v4-27b-it (SEA-language focused, secondary for coding).
- OpenCode Zen (+1): deepseek-v4-flash-free returned to the free pool (re-added with a re-verify note since it is still absent from the docs pricing table).
- LLM7 (+1): GLM-5.3-Flash is free again (turbo tier, verified with a live unauthenticated chat probe).
- SiliconFlow (+1): XingChenAGI/Xing4.0-29B, a new $0 engineering-focused model.
Fixed
- NVIDIA key testing was silently broken in production:
PROVIDER_TEST_MODEL_OVERRIDES.nvidiaprobed keys with the model NVIDIA killed today, so a valid key would have been reported dead. It now probes with moonshotai/kimi-k3, then gpt-oss-120b. - Comment/value mismatches from the 09-15 audit: OpenRouter glm-5.2:free claimed a 32k ctx fix that was never applied to the value (now 32k for real: the free endpoint caps at 32768); SiliconFlow GLM-Z1-9B and Codestral carried false "ctx fixed to..." comments contradicting correct values (comments removed, values untouched).
- NVIDIA ctx corrections: gemma-4-31b-it, diffusiongemma-26b-a4b-it and nemotron-3-nano-omni are 262k (official contextLength 262144), not 256k. Kimi K3 re-tiered to S+ with its 76.8% SWE-bench Verified score.
- Pollinations aliases upgraded upstream: laguna now serves Laguna S 2.1 (was XS.2), deepseek serves DeepSeek V4 Flash (was V3), kimi-code serves Kimi K2.7 Code, qwen-coder serves Qwen3 Coder 30B (re-scored to its own 51.6%), and openai resolves to GPT-5.4 Nano (re-tiered to the nano class). Labels and scores now match what actually answers. Header updated: /v1/chat/completions now requires a free API key (401 anonymous).
- OVHcloud embedding models: bge-m3 and bge-multilingual-gemma2 got their real 8k context.
- OrcaRouter: the in-file comment claimed the fusion family is pay-as-you-go while the model list said otherwise; the comment is corrected and the undocumented fusion models stay out until their free status is confirmed.
- Tests: router failover, endpoint-installer and key-discovery suites pinned dead NVIDIA ids; all swapped to live ids (1154/1154 tests pass).
- Docs: README and docs/providers.md refreshed to 24 providers / 256 models (regenerated from sources.js), website catalog copy synced.