Free LLM APIs in 2026: Rate Limits and Training Terms

Short answer: OpenRouter gives the most daily requests, 1,000 a day once you have bought $10 of credit. Cerebras gives the largest token budget, 1M tokens a day on GPT OSS 120B, but its $5 trial credit expires 30 days after activation. When free access ends, Mistral Small 4 is the cheapest paid route we verified, at $0.15 input and $0.60 output per 1M tokens.

Free tiers change often, and each vendor states its caps in a different place. Apart from the two labelled counts below, this page uses only what each vendor says on its own pricing, rate-limit or changelog page, checked on 10 October 2026. Where a cap sits behind a sign-in, the table says so instead of guessing.

Devshot data: 22 of the 659 models in OpenRouter's public model list carry the :free suffix today, counted from its models API on 10 October 2026. Toolradar data: the AI model deployment ranking evaluated 42 tools.

Free tiers at a glance

OpenRouter: the most daily requests, once you buy $10

Free models on OpenRouter are the ones whose ID ends in :free. Every account gets 20 requests a minute on them. An account with no credit gets 50 requests a day. Buying at least $10 of credit lifts that ceiling to 1,000 a day. Failed attempts count against the daily cap, and a negative balance blocks free models until you add credit.

Buying credit carries a fee. OpenRouter charges 5.5% on credit purchases, with a $0.80 minimum on card payments. On a $10 top-up, 5.5% would be $0.55, so the $0.80 minimum applies. Inference carries no markup of its own: the FAQ says it is billed at the provider's list price on every plan. Paid usage therefore costs whatever the model you route to lists at, and the cost changes by model.

The cap counts requests, not tokens, and OpenRouter publishes no token cap for free models. A 20-token prompt and a 20,000-token prompt both use one request. For a gateway that adds its own fee on top of model prices, see the OpenRouter alternatives ranking, and the best API gateways list covers the wider field. That fee is a separate cost from the free tier here. The OpenRouter entry on Toolradar has the directory details.

OpenRouter says prompts are not logged by default. It skips providers that log prompts, or whose policy it cannot confirm, unless you switch on the model training toggle in your privacy settings. That is a routing rule, so check the policy of the specific model you pick before you send private text.

Cerebras: the largest token budget, on a 30-day clock

Cerebras has no permanent free tier. New accounts get a one-time $5 Free Trial credit after they add a valid payment method. The credit expires 30 days after activation. Cerebras says you are not charged or enrolled in paid usage automatically, so access stops when the credit runs out until you buy more.

While the trial lasts, GPT OSS 120B and Qwen 3.8 27B each allow 5 requests a minute, 30,000 uncached tokens a minute (90,000 total), and 1 million tokens an hour and a day. The per-minute request cap is the tighter limit. A chat app with a few users at the same time hits 5 requests a minute before the daily token cap matters.

The token budget is generous, but the credit is not. At list price, GPT OSS 120B costs $0.35 per 1M input tokens and $0.75 per 1M output tokens on Cerebras's developer tier. A workload of 1M tokens a day, split evenly, costs $0.55 a day, so the $5 credit covers about nine days of it. Qwen 3.8 27B lists at $0.99 input and $1.49 output, which puts it at more than twice the cost of GPT OSS 120B on the same mix.

Cerebras's pricing page does not say whether prompts from free accounts train its models. Read its data terms before you run customer text through the trial.

Google Gemini API: four free models, caps behind a sign-in

Four current models are free of charge on the Standard tier: Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.1 Flash-Lite. The Gemini model on Toolradar is the 3.1 Flash-Lite listing.

Google publishes per-model free caps only in Google AI Studio, after sign-in. Its rate-limit docs explain the rules: limits apply per project, and daily quotas reset at midnight Pacific time. The numbers themselves are not in the docs, so this page gives no requests-per-minute figure for Gemini.

The training line matters more than the caps. Google's pricing page says content on the Free plan is used to improve its products, and the Paid plan says it is not. For a prototype that handles real user text, that single line decides the route. Gemini is the only provider in this list whose pricing page states the paid-tier rule in writing.

On price, Gemini 3.1 Flash-Lite is the cheapest Gemini model at $0.25 input (text, image and video; $0.50 for audio) and $1.50 output per 1M tokens. Gemini 3.8 Flash and 3.6 Flash cost $0.75 input and $3.75 output through December 31, 2026. From January 1, 2027, both double to $1.50 and $7.50. A team that builds on 3.8 Flash now should budget for that step.

Mistral: free mode with caps behind the console

Mistral's free mode lets you create API keys and use "included monthly usage within the limits shown on the Limits page" in the admin panel. The public docs give no requests-per-second, tokens-per-minute or monthly figure, so this page cannot quote one. Accounts move up tiers by billed usage, not by prepaid credit. Tier 2 starts after more than $20 billed, Tier 3 after $100, and Tier 4 after $500. Prepaid credit does not count toward those thresholds. Mistral AI on Toolradar lists the vendor record.

Mistral Small 4 is the cheapest paid model in this list on both input and output. Cached input drops to $0.015 per 1M tokens, and batch processing costs half. The price table is in dollars and gives no effective date, so re-check it before you budget.

Mistral's plan table lists model training as "Opt-out" and does not say which plans that covers. Read the data terms before you send customer text through the free mode.

Cohere trial keys: good for testing, not for production

Cohere's trial keys are free, with 20 requests a minute on chat models and a cap of 1,000 API calls a month. Cohere says trial keys cannot be used for production or commercial purposes, so a free key is for evaluation only. Its pricing page lists no current price for Command A, so this page quotes none. Cohere on Toolradar is the directory record.

GitHub Models is gone

GitHub Models was retired on July 30, 2026. Its playground, model catalog, inference API and bring-your-own-key option no longer work for any customer. GitHub points people to Microsoft Foundry for model access and to GitHub Copilot instead. Any list that still names GitHub Models as a free route is out of date.

What a prototype costs after the free tier

Free caps are for testing. To compare real costs once free access ends, this page prices one scenario: 1 million tokens a day for 30 days, half input and half output, at each route's first paid list price. For Mistral Small 4, half the tokens at the input price plus half at the output price comes to $0.375 a day, and $0.375 x 30 = $11.25 a month.

Mistral Small 4 is the cheapest route at any input share, because it is lower on both input and output. Gemini 3.1 Flash-Lite has the cheaper input, so it beats Cerebras GPT OSS 120B only when input is more than about 88% of the tokens. At the even split used here, Gemini costs $9.75 more over 30 days. OpenRouter is not in the table: it bills the model you choose at that model's list price, so its cost depends on the model.

Which free route to use

For many daily calls with cheap models, OpenRouter is the route. Top up once to raise the daily cap, and accept that paid usage then follows each model's list price. For a short test measured in tokens, Cerebras's trial gives the most volume, but at a 1M-token-a-day workload the credit runs out in about nine days. For Google models on the free tier, use them only when training on your text is acceptable, and move to the paid tier when it is not.

For a steady workload you will keep paying for, the cheapest scenario here is Mistral Small 4. Choose the model on your own test set first, because this ranking says nothing about quality. For the coding side of the bill, see AI for developers, Claude Code pricing, Cursor's plans, the best AI code generators and AI-assisted coding. Coding assistants with a free plan, and the cap each one publishes, are in Best Free AI Coding Tools in 2026.

How this page was checked

• Each price, cap and date comes from a vendor's own pricing, rate-limit or changelog page, read on 10 October 2026. Gemini's pricing page was last updated on 9 October 2026 (UTC).
• The free-model count comes from OpenRouter's public models API, and the tool count from Toolradar's ranking, both read on the same day.
• A cap appears only when the vendor states it. Caps behind a sign-in are named as such, with no number.
• The cost scenario is 1 million tokens a day for 30 days, half input and half output, at first paid list prices, with no discount and no caching. It follows the per-million-unit pricing used in the text to speech API ranking. Another cost comparison on this site is object storage pricing compared.
• Rankings are by cost for that scenario only. They do not measure model quality.

Frequently asked questions

Which free LLM API has the most requests per day?

OpenRouter. Its free models allow 50 requests a day with no credit, and 1,000 a day after you buy $10 of credit. Every account gets 20 requests a minute on them.

Is there a free LLM API with a token limit?

Cerebras. Its Free Trial allows 1 million tokens an hour and a day on GPT OSS 120B and Qwen 3.8 27B, but the $5 credit expires 30 days after activation.

Do free LLM APIs train on my prompts?

It depends on the vendor. Gemini's free plan uses content to improve Google's products, and its paid plan does not. OpenRouter does not log prompts by default and skips providers that log them unless you switch on the training toggle. Cerebras does not say on its pricing page. Mistral's plan table lists "Opt-out" without naming the plans it covers.

Can I run a product on a free LLM API trial key?

Not with Cohere. Its trial keys are capped at 1,000 API calls a month and cannot be used for production or commercial purposes.

What does an LLM API cost after the free tier?

For 1 million tokens a day over 30 days, split evenly between input and output, Mistral Small 4 costs $11.25 a month at its first paid list price, the lowest in this comparison. Gemini 3.1 Flash-Lite costs $26.25 on the same mix.

Is GitHub Models still a free option?

No. GitHub retired GitHub Models on July 30, 2026, and points users to Microsoft Foundry and GitHub Copilot instead.

Cite this: Devshot, "Free LLM APIs in 2026: Rate Limits and Training Terms", October 2026.

Some offers on this page may be paid placements or contain affiliate links.