Free LLM API: every genuinely free option, with the catch for each

The comparison, with real numbers

Every figure below was read from the provider’s own documentation in August 2026. Where a vendor has stopped publishing numbers, this table says so rather than repeating a figure from a blog post.

OpenRouter free models

OpenRouter aggregates hundreds of models behind one OpenAI-compatible endpoint, and a subset carry a :free suffix. The rate limits on those are documented precisely, which is more than most of this list manages.

The threshold is worth reading carefully: it is a lifetime condition, not a balance. Once you have ever bought 10 credits, your free-model daily cap is 1,000 requests, and it stays there even after you have spent that balance. Ten dollars is the cheapest twentyfold increase in this entire guide.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<vendor>/<model>:free",
    "messages": [{"role": "user", "content": "Explain a git worktree in two sentences."}]
  }'

Google Gemini: free, with a real condition

Gemini has the most generous free offering here by capability. The entire Flash and Flash-Lite line is listed as "free of charge" for standard input and output: Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, and 2.5 Flash and Flash-Lite. (Gemini 2.0 Flash was shut down on 1 June 2026.) The Pro models are not included.

The second thing to know is that Google stopped publishing free-tier RPM and RPD figures in its rate-limits documentation. That page now says limits "can be viewed in Google AI Studio" and points at the dashboard. Any specific requests-per-minute number you find for the Gemini free tier in a blog post is therefore unverifiable against a primary source, including numbers that were correct when written. Check your own account.

Groq and Cerebras: fast, open-weight, small caps

Both sell inference speed on custom silicon, and both let you try it free with open-weight models. They are the best options in this guide when latency is what you are testing.

The Groq numbers hide a trap worth doing the arithmetic on: 1,000 requests a day sounds generous, but the 200,000 tokens a day ceiling means an average of 200 tokens per request before you hit the wall. For a chat toy that is fine. For a coding agent, where a single turn routinely carries tens of thousands of tokens of context, the daily budget is closer to a handful of turns.

Local models: the zero-cost floor

Running the weights on your own machine is the only entry here with no rate limit, no daily cap, no key, and no question about who reads your prompts. Ollama exposes an OpenAI-compatible endpoint on localhost, so most tooling points at it with one environment variable.

ollama pull gpt-oss:20b
ollama serve

# then point any OpenAI-compatible client at it
export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_API_KEY=ollama

The cost moved rather than disappeared. gpt-oss-20b and its 20B-class peers run acceptably on a 16GB machine. The 120B-class models, gpt-oss-120b included, want an 80GB-class GPU or a large unified-memory Mac, and on CPU alone they will run but slowly enough that you will stop using them. Weigh that against a paid API before buying hardware: for most individuals the hardware costs more than years of the API would.

Credits, and the churn problem

Beyond permanent free tiers there are one-off credit pools, which are worth taking and dangerous to plan around.

• Signup credits. Cerebras gives $5. Most inference vendors give something similar for a new account. Fine for evaluation, gone in a week of real use.
• Startup programmes. Most major labs and clouds run one, typically gated on an incorporated company and sometimes on an investor introduction. The amounts are large enough to matter and the applications take weeks.
• Student and education access. Several vendors discount or waive fees with a verified .edu address. Terms change every academic year.
• Cloudflare Workers AI. 10,000 Neurons a day free, then $0.011 per 1,000 Neurons. Neurons are a normalised compute unit rather than tokens, so the effective free allowance depends heavily on which model you pick.

For the tier above free, where you are paying but want the cheapest workable rate, see the cheapest LLM APIs and inference providers compared.

Questions people ask

Keep reading

Continue through Inference providers with Inference providers: the honest 2026 comparison.