Free LLM APIs in 2026: Every Free Tier and Its Limits

Eight providers give some form of free API access as of October 2026: Google (Gemini), Mistral, Groq, OpenRouter, Cloudflare, GitHub, Cerebras and OpenAI. Two more give trial allowances for embeddings or rerank (Voyage AI, Cohere). Anthropic and DeepSeek have no free tier, and Hugging Face gives free accounts no inference credits. Every number below was read on 30 October 2026 from a vendor page linked in the Sources section. Where a page does not print a number, the table says so instead of filling the gap.

Provider by provider

Google Gemini API

The free tier covers the Flash and Flash-Lite lines plus Gemini 2.5 Pro. Gemini 4 Argon, 3.1 Pro Preview and the image models are marked "Not available" on it. The Flash-Lite free tier has no context caching, and Batch works only with billing enabled.

The rate-limits page does not print the per-model numbers. It says they "can be viewed in Google AI Studio" and that requests per day reset at midnight Pacific. Limits apply per project rather than per key. The free tier has no billing cap because there is nothing to bill.

The pricing page marks free-tier content as "used to improve products: Yes". On the paid tier the same field says No. This is the clearest written cost of the Gemini free tier.

Google has also paused the free tier before, when demand spiked. The post below is from May 2025. Google switched off free 2.5 Pro access in the API then, while keeping it on in AI Studio.

Mistral

Free mode is on by default in La Plateforme (AI Studio). The limits are 1 request per second, 500,000 tokens per minute and 1 billion tokens per month, counted per workspace. The help article says the limits apply to "all models that are compatible with the Free-tier" without listing them.

The monthly token budget is the most generous printed number in this table. The 1 request per second ceiling is the binding limit for anything interactive with more than one user. Neither fetched page says whether free-mode prompts are used for training.

Groq

Groq has a Free plan, but the rate-limits page prints one set of base limits and does not say which column applies to Free. For gpt-oss-120b and qwen3.8-27b those limits are 30 requests per minute, 1,000 per day, 8,000 tokens per minute and 200,000 tokens per day. Whisper is 20 requests per minute and 2,000 per day.

The 8,000 tokens-per-minute line matters more than the request count. One request with a 6,000-token prompt uses most of a minute's budget. Batch and Flex processing require the Developer plan.

OpenRouter

Models with a :free suffix cost nothing. Everyone gets 20 requests per minute. The daily cap depends on purchase history: 50 requests per day if you have bought fewer than 10 credits, 1,000 per day once you have bought 10 or more. A negative balance blocks the free models.

The openrouter/free router picks only among free models. The :free entry for openai/gpt-oss-120b keeps the 131K context. The fetched pages do not say whether free-model traffic is used for training. That depends on the upstream provider serving the request.

Cloudflare Workers AI

Workers AI gives 10,000 Neurons per day on both the Workers Free and Workers Paid plans. Above that, Paid accounts pay $0.011 per 1,000 Neurons. A Neuron is Cloudflare's compute unit, so the same 10,000 buys a different number of tokens for Llama 3.1 8B than for Llama 3.3 70B. Check the per-model Neuron rate on the pricing page before planning around the daily number.

GitHub Models

GitHub Models is in public preview and free with GitHub Free and Copilot Individual accounts. Low-tier models get 15 requests per minute, 150 per day, 8,000 input and 4,000 output tokens per request and 5 concurrent requests. High-tier models get 10 per minute, 50 per day and 2 concurrent. Embeddings get 15 per minute, 150 per day and 64,000 tokens. Business and Enterprise plans get higher limits.

The 8,000-token input cap rules out long documents. Production use goes through a paid Azure token. The official docs page was not reachable from our fetch environment. The numbers come from a mirror of the same Markdown source.

Cerebras

The Free Trial covers gpt-oss-120b and qwen-3.8-27b. Its limits are 5 requests per minute, 30,000 uncached tokens per minute, 90,000 total tokens per minute, 1M tokens per hour and 1M tokens per day. The Developer tier removes the hourly and daily caps and lifts gpt-oss-120b to 1,000 requests per minute and 1M uncached tokens per minute. Five requests per minute is enough for a demo but too few for a shared tool.

OpenAI

Two things are easy to confuse here. The usage-tiers page lists a "Free" tier with a $100 monthly usage limit and an eligibility rule (located in an allowed geography). It prints no free tokens, RPM or TPM for it.

The separate data-sharing programme is where free tokens come from. Organisations that opt in to sharing their prompts and completions get up to 1M complimentary tokens per day on the larger model group. The allowance on the smaller group is 10M per day, and lower usage tiers receive 250K and 2.5M. The counter resets at 00:00 UTC and overage is billed at normal rates. Fine-tuned models, fine-tuning, evals and tool use are excluded. Zero Data Retention and Enterprise orgs cannot opt in, and the account needs a positive balance.

Two caveats apply. First, the page says OpenAI will give 30 days' notice before ending the programme, so it can end. Second, its model list (GPT-5, GPT-4.1, GPT-4o, o-series) and tier names predate the current GPT-6 lineup. The quota figures may be stale until the page is updated.

Voyage AI and Cohere

Voyage gives 200M free tokens per account for the voyage-4 embedding family and rerank-3, and 50M for the older domain models. The pricing page describes the allowance as "the first 200 million tokens" and mentions no monthly reset, so treat it as one-off. Cohere trial keys are free and rate limited, with no numbers on the pricing page and an explicit ban on production or commercial use.

Free requests per day, side by side

Only three providers publish a daily request cap: Groq, OpenRouter and GitHub Models. Gemini gives no figure, Mistral caps per second and per month, Cloudflare counts Neurons, and Cerebras caps tokens. The second chart shows the daily token caps that do appear on the pages.

What free costs you

Your data

Google states in writing that free-tier content is used to improve products and paid-tier content is not. OpenAI's token programme openly trades your traffic for tokens. Mistral, Groq, OpenRouter and Cloudflare do not state a training policy on the pages fetched for this article. You have to read their terms yourself before sending customer data through them.

Rate limits that rule out production

Five requests per minute (Cerebras) or 1 per second (Mistral) is fine for one developer. It fails the first time two users press a button at once. Caps of 50 to 1,000 requests reset once a day, so a burst at 9 a.m. can lock the tool until the next reset. A single long prompt can hit a token-per-minute cap such as Groq's 8,000.

No SLA and no guarantee the tier survives

The 2025 Gemini 2.5 Pro pause above shows the pattern. A vendor can switch a free model off when paid demand needs the capacity. OpenAI's programme reserves the right to end with 30 days' notice. GitHub Models is a public preview.

Model turnover

Free tiers tend to carry the small and previous-generation models. Gemini 2.5 Pro is on the free list; Gemini 4 Argon is not. Groq's and Cerebras's free models are open-weight models you could also run elsewhere. When the vendor retires a model, your free pipeline retires with it.

Where free tiers fit

Free tiers suit four kinds of work:

• prototyping before a budget exists;
• running evals where you control the pace;
• student and classroom projects;
• internal tools with a handful of users and no uptime promise.

Anything customer-facing should start on a paid key, even if the paid bill stays near zero.

When to switch to paid

Take a small production load: 1,000 requests a day, each with a 1,500-token prompt and a 300-token answer. That is 45M input and 9M output tokens a month. The table shows what it costs on a selection of low-priced paid models with verified prices.

The same 1,000 requests a day sit exactly on Groq's printed daily request cap. On tokens, though, they need 1.8M a day against Groq's 200,000-token daily cap. OpenRouter would need a purchase of 10 credits. The load exceeds GitHub Models' limits roughly 7 to 20 times over, and Cerebras's 1M tokens per day falls short of the 1.8M needed. Mistral's free mode would carry the load on tokens (54M of a 1B monthly budget). Its 1 request per second ceiling, however, makes any two overlapping users queue.

Money is a minor factor in the decision. The paid bill for this load is between $2.70 and $12.15 a month, less than an hour of engineering time spent working around rate limits. Switch when the tool has more than one concurrent user, when it touches customer data, or when a failed call costs you anything. LLM API pricing, October 2026 has the full price table, and the cheapest LLM APIs by task has the cheapest picks per task.

What to do next

• Prototype on Gemini Flash, Groq or OpenRouter :free, and check the training policy before sending anything you would not post publicly.
• Record the printed limits (RPM, RPD, TPM) for the provider you pick, and add a retry with backoff on 429 responses from day one.
• Move customer-facing traffic to a paid key on a $0.10-per-million-input model; at this prompt size, budget about $10 a month per 1,000 daily requests.
• Keep the free key for evals and regression runs, where you control the pace.
• If your bill is already past the point where free tiers matter, we audit pipelines for the usual leaks; see an LLM cost audit.

Sources

• Gemini API pricing
• Gemini API rate limits
• Mistral: What are the limits of the free tier
• Groq rate limits
• OpenRouter API limits
• Cloudflare Workers AI pricing
• GitHub Models: Prototyping with AI models (mirror of the GitHub Docs source; official page not reachable on 30 Oct 2026)
• Cerebras rate limits
• OpenAI usage tiers
• OpenAI: Free tokens for sharing data
• Voyage AI pricing
• Cohere pricing
• Hugging Face Inference Providers pricing
• Anthropic rate limits
• DeepSeek pricing
• OpenRouter: gpt-oss-120b
• OpenAI models
• Claude API pricing
• Mistral inference pricing
• Groq: gpt-oss-120b model page