You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Short answer: 24 of the 26 providers in this dataset are OpenAI compatible — every one except IBM watsonx.ai. So in most cases you change two things, base_url and the key, and your existing OpenAI client keeps working.
Here are the base URLs, all re-read on 2026-07-25.
Cloudflare is the one that is not a drop-in: the account ID is part of the path, so it cannot be pasted into a client that only accepts a host.
Also compatible, but not a permanent free tier
OpenRouter (https://openrouter.ai/api/v1), Cerebras (https://api.cerebras.ai/v1), Vercel AI Gateway (https://ai-gateway.vercel.sh/v1), Together (https://api.together.xyz/v1), Nebius (https://api.tokenfactory.nebius.com/v1), Perplexity (https://api.perplexity.ai), DeepInfra (https://api.deepinfra.com/v1/openai), Chutes (https://llm.chutes.ai/v1), Scaleway (https://api.scaleway.ai/v1).
Pointing a coding agent at one
Most agents read OPENAI_BASE_URL and OPENAI_API_KEY from the environment, or take the equivalent two fields in a settings file. Set the base URL to a row above, set the key to one you generated yourself, and pick a model the provider actually serves — a wrong model ID looks exactly like a broken endpoint. Per-client setup, written out: docs/clients.md
For a coding agent specifically, the daily ceiling usually matters more than the per-minute one, because agents burst. Groq's 1,000/day on the 70b and 14,400/day on the 8b are the largest published day figures in the free set.
What "OpenAI compatible" does not mean here
It means the endpoint speaks the chat-completions shape. It does not promise every parameter is honoured, that tool calling behaves identically, or that streaming has the same semantics. It also says nothing about the model being any good. This dataset records compatibility as the provider documents it, with the date it was checked, and nothing more.
One thing that would help: reply with the client you are using and whether the base URL above worked unchanged. The compatible/not-compatible flag is per provider today; the interesting failures are per client, and I can only see those if someone hits one and says so.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Short answer: 24 of the 26 providers in this dataset are OpenAI compatible — every one except IBM watsonx.ai. So in most cases you change two things,
base_urland the key, and your existing OpenAI client keeps working.Here are the base URLs, all re-read on 2026-07-25.
Free tiers, no credit card
https://api.groq.com/openai/v1https://generativelanguage.googleapis.com/v1beta/openai/https://api.siliconflow.com/v1https://api.sambanova.ai/v1https://api.cohere.ai/compatibility/v1https://api.fireworks.ai/inference/v1https://api.mistral.ai/v1https://api.z.ai/api/paas/v4https://api.moonshot.ai/v1https://dashscope-intl.aliyuncs.com/compatible-mode/v1https://router.huggingface.co/v1https://api.novita.ai/openaihttps://ollama.com/v1https://text.pollinations.ai/openaihttps://api.cloudflare.com/client/v4/accounts/ACCOUNT_ID/ai/v1Cloudflare is the one that is not a drop-in: the account ID is part of the path, so it cannot be pasted into a client that only accepts a host.
Also compatible, but not a permanent free tier
OpenRouter (
https://openrouter.ai/api/v1), Cerebras (https://api.cerebras.ai/v1), Vercel AI Gateway (https://ai-gateway.vercel.sh/v1), Together (https://api.together.xyz/v1), Nebius (https://api.tokenfactory.nebius.com/v1), Perplexity (https://api.perplexity.ai), DeepInfra (https://api.deepinfra.com/v1/openai), Chutes (https://llm.chutes.ai/v1), Scaleway (https://api.scaleway.ai/v1).Pointing a coding agent at one
Most agents read
OPENAI_BASE_URLandOPENAI_API_KEYfrom the environment, or take the equivalent two fields in a settings file. Set the base URL to a row above, set the key to one you generated yourself, and pick a model the provider actually serves — a wrong model ID looks exactly like a broken endpoint. Per-client setup, written out: docs/clients.mdFor a coding agent specifically, the daily ceiling usually matters more than the per-minute one, because agents burst. Groq's 1,000/day on the 70b and 14,400/day on the 8b are the largest published day figures in the free set.
What "OpenAI compatible" does not mean here
It means the endpoint speaks the chat-completions shape. It does not promise every parameter is honoured, that tool calling behaves identically, or that streaming has the same semantics. It also says nothing about the model being any good. This dataset records compatibility as the provider documents it, with the date it was checked, and nothing more.
Filterable directory of all 26: https://xyzs996.github.io/free-llm-api/ · check a key from your browser: https://xyzs996.github.io/free-llm-api/verify.html
One thing that would help: reply with the client you are using and whether the base URL above worked unchanged. The compatible/not-compatible flag is per provider today; the interesting failures are per client, and I can only see those if someone hits one and says so.
All reactions