OpenRouter Free Models: Limits, Credits and API Setup

OpenRouter is one API in front of hundreds of models. Some of them have a free variant whose ID ends in :free: you call it with your normal key and pay $0 per token, within a shared daily cap. Every number below was checked on OpenRouter's docs, pricing page and live models API on 2026-10-21.

What is free right now

There are two counts, and both are right. On 2026-10-21 the free models filter and the models API with ?output_modalities=all listed 24 :free model IDs. The plain models API listed 17, because by default it only returns models that output text. By type that day:

Text chat examples from that day:

The free embedding models were nvidia/nemotron-3-embed-1b:free (32,768-token context), nvidia/llama-nemotron-embed-vl-1b-v2:free (text and image input, 131,072) and liquid/lfm-2.5-embedding-350m:free (512). The reranker was nvidia/llama-nemotron-rerank-vl-1b-v2:free, and the text-to-speech model fish-audio/s2.1-pro-free:free.

The lineup changes often, so do not hard-code it: the current list is on the free models filter or in the models API (filter IDs ending in :free; add ?output_modalities=all to include embedding, rerank and speech models). Two notes:

• openrouter/free, the Free Models Router, "automatically selects a free model at random from the available free models" and filters for the features your request needs (images, tool calling, structured outputs). The response's model field says which one answered; you cannot pick it.
• Not every $0 price is free: the Lyria music models show a token price of 0 but charge "$0.08 per song".

No DeepSeek, Llama or Qwen model had a :free variant on the check day, even though an older OpenRouter page still lists them as examples. For free DeepSeek elsewhere see the free DeepSeek API guide.

Limits: 50 or 1,000 requests a day

From OpenRouter's limits page, for any model ID ending in :free (the page makes no exception for embedding or rerank models, so count them against the same cap):

• The daily count resets each UTC day. "Making additional accounts or API keys will not affect your rate limits."
• "Failed attempts count toward that daily cap, and a negative balance blocks free models until you add credits" (pricing). A negative balance returns HTTP 402 even on :free models.
• Check what you have left with GET https://openrouter.ai/api/v1/key: is_free_tier and free_model_daily_requests (used, limit, remaining). The per-minute limit is not reported there.
• A 429 can come from OpenRouter (your cap) or from the upstream provider (the free endpoint is busy). Only OpenRouter's own 429 carries X-RateLimit-* headers.

To see how far 50 or 1,000 requests a day go for your workload, put the numbers in the rate limit calculator.

When to buy credits

Buying credits is the only way to lift the daily cap, and you only have to do it once: the threshold counts credits purchased "all time". It is worth it when 50 requests a day is not enough (an agent can use that in one task) or when you also want paid models through the same key.

• Payment by card, AliPay or crypto (USDC). The card fee is 5.5% ($0.80 minimum) and crypto 5% (FAQ); there is no minimum spend on the Standard plan.
• OpenRouter "reserve[s] the right to expire unused credits after one year of purchase".
• Free requests do not spend credits; keep the balance above zero so free models keep working.

Call a free model

1. Sign up at openrouter.ai and create an API key in your account settings.
2. Call the OpenAI-compatible endpoint: curl https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "google/gemma-4-31b-it:free", "messages": [{"role": "user", "content": "Hello"}]}'
3. For embeddings, call curl https://openrouter.ai/api/v1/embeddings -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "nvidia/nemotron-3-embed-1b:free", "input": "Hello"}'. Pass an array as input to embed several texts in one request, which also saves requests against the daily cap.
4. With the OpenAI SDK, set the base URL to https://openrouter.ai/api/v1 and the model to any :free ID or openrouter/free. The optional HTTP-Referer and X-OpenRouter-Title headers credit your app in OpenRouter's rankings.

If you set a models fallback list, keep every entry free: a fallback to a paid model "incurs charges even though the first model was :free". To use these models in a coding agent, the free AI coding agent guide walks through the Cline setup.

Privacy settings decide which free models you can reach

OpenRouter itself does not log prompts by default, but the providers behind free endpoints have their own policies. On the check day, 11 of the 17 text :free endpoints were marked as providers that may train on your data (all seven Nemotron models, both Inkling models, Laguna S and XS 2.1, LFM2.5). OpenRouter has a separate training setting for free models in your privacy settings; with it switched off, "OpenRouter will not route to providers that train", so those 11 fail to route. Turn it on only for prompts you are fine sharing.

Limits and the catch

• Free endpoints are single-provider and can be down. Each :free model is served by one provider; on 2026-10-21 Nemotron 3 Nano Omni showed 8.2% uptime over the last day. Keep a second free model or openrouter/free as a fallback.
• Not for production. OpenRouter's FAQ says free models "are usually not suitable for production use".
• The lineup rotates. Read model IDs from config, and re-check the free filter before you build on one model.
• Other free LLM APIs with their own keys: see the free Groq API key, free Gemini API key and free OpenAI API access guides, or let the free LLM API picker choose one by context size, rate limit and card requirement.