The best free LLM APIs in 2026: a live-probed list

Last updated: July 6, 2026

TL;DR

A free LLM API lets you call a hosted language model without paying. Across 16 providers there are about 110 free models right now, but only ~57 are genuinely free with permissive terms. The rest are eval-only, capped, anonymous-access, or off-limits under their own terms of service. This list sorts them by what's actually usable, refreshed from a live-probed catalog.

How many free LLM APIs are there really?

There are roughly 110 free models across 16 providers as of June 2026, but the honest number is smaller: about 57 are truly free with terms that allow personal use. Most "free LLM API" lists are just stale link dumps, every tier counted the same. The thing nobody separates out is that "free tier" means four pretty different things, and three of them come with strings attached.

That's the split this list uses. The counts come from a catalog that gets re-probed, which means dead tiers get culled and the number drifts over time:

One warning up front: free tiers move around weekly. Take any specific limit below as a starting point, not gospel, and check the provider's docs before you lean on it.

The master list

Every provider, what you get, and the catch. The models listed are examples, not the whole menu.

One more worth knowing: OVH AI Endpoints documents anonymous access (around 2 requests per minute per model) with no signup. It drifts in and out of the catalog, but it's a genuinely handy no-account option.

Tier 1: truly free, start here

These have a real free tier and terms that let you use it for personal stuff. If you just want to call a free model today without spending an hour in someone's legal page, start here.

• Groq is the standout for speed. The free tier runs Llama 3.3, Llama 4, GPT-OSS, and Qwen3 fast, and the agreement is fine with you building it into your own app.
• Cerebras runs big models like Qwen3 235B surprisingly fast on the free tier. The one rule: don't sell or hand off your keys.
• Mistral has a free experimentation tier across Large 3, Codestral, and Devstral, and it's allowed for personal and internal business use.
• OpenRouter is where you get the most free routes in one place, roughly two dozen tagged :free. Availability rotates, so don't pin your hopes on one specific model.
• Ollama Cloud gives you free cloud access to GLM-4.7, Kimi K2, gpt-oss, and Qwen3, with one request at a time and 5-hour session caps.
• Z.ai / Zhipu has a non-commercial research carve-out for GLM-4.5 and GLM-4.7 Flash.

Tier 2: free, but read the limits

Free to call, with a real asterisk: either a hard rate cap, or terms that say "evaluation" or "prototyping" out loud.

• NVIDIA NIM has a huge catalog, but the free access is evaluation-only and runs at 40 requests a minute. Great for kicking the tires, not for a backend.
• GitHub Models gives you GPT-4.1 and GPT-4o free, but the terms scope it to experimentation and prototyping.
• Google Gemini has a strong free tier (Gemini 2.5 Flash, 3.x previews), but a 2026 terms change pointed it toward business rather than consumer use. A self-hosted dev setup is still defensible; just know the clause is there.
• Cloudflare Workers AI hands you a daily allocation across Kimi K2, GLM-4.7, GPT-OSS, and Granite 4, with no anti-proxy clause but only general subscription terms.

Tier 3: no account needed

The fastest path to a free completion: no signup, no key. Quality and uptime bounce around, but for a quick test or a throwaway script, nothing beats them.

• Kilo has :free routes that work anonymously.
• Pollinations serves GPT-OSS 20B with no account.
• LLM7 gives you GPT-OSS, Llama 3.1, and GLM anonymously.
• OpenCode Zen runs promo anonymous access to DeepSeek V4 Flash and Nemotron.
• OVH AI Endpoints documents anonymous access at about 2 requests a minute per model.

Tier 4: skip these (for personal use)

Free, but the terms forbid what most people actually want to do. Cohere's free trial of Command R+ and Command-A still bans "personal, family, or household purposes." The models are good; the license just isn't built for a hobby project. Read it before you build on it.

What you actually get

Free doesn't mean equal. Past the model name, three things decide whether a free tier is any use to you:

• Rate limits and token budgets. A 40-a-minute eval tier and a generous experimentation tier are both "free" and they feel nothing alike when you're actually using them.
• Context window. Some free models give you a few thousand tokens, others give you six figures. If you're feeding in long files, that's the whole game.
• Tool calling and vision. About 79 of these free models do tool calling and 15 do vision. If your app needs function calls, filter to the tool-capable ones before you fall in love with a model that can't.

That's the reason for a live-probed catalog instead of a static list. A model that quietly stops returning tool calls gets flagged and dropped, instead of sitting there looking fine.

How to use all of them at once

One free tier is a toy. Stack them and they're a real working pool, except now you're juggling 16 SDKs and 16 rate limits, which is its own misery. The open-source FreeLLMAPI project (by Tashfeen Ahmed, MIT) puts the free tiers behind one OpenAI-compatible endpoint, with a router that fails over when a provider taps out.

The move that actually changes how you use these, though, is keeping a smart model in charge and letting it hand the cheap work to the free pool. That's the whole idea in how to put Claude Code in charge of a free LLM API pool: Claude orchestrates, the free models do the grunt work, and a rubric decides what's worth spending real tokens on.

Frequently asked questions

Is there a truly free LLM API?

Yes. Several providers offer genuinely free tiers, including Groq, Cerebras, Mistral, OpenRouter, Ollama Cloud, and Google Gemini. About 57 free models across these permissive-terms providers are usable for personal projects today, though specific rate limits and model availability change often.

What is the best free LLM API?

For speed and a permissive free tier, Groq is the standout. For the widest selection in one place, OpenRouter aggregates the most free routes. For no signup at all, Kilo, Pollinations, and LLM7 offer anonymous access. The best one depends on whether you optimize for speed, selection, or zero friction.

Is there a free LLM API with no account or key?

Yes. Kilo, Pollinations, LLM7, OpenCode Zen, and OVH AI Endpoints all offer anonymous access with no signup. Uptime and quality vary more than account-based tiers, but for a quick test or a throwaway script they're the fastest path to a free completion.

Are free LLM APIs legal to use?

For personal experimentation, most are fine, but the terms vary by provider and some are restrictive. Cohere's free tier forbids personal use, NVIDIA's is evaluation-only, and GitHub Models is scoped to prototyping. Use one account per provider, don't resell access, and read each provider's terms. This is informational, not legal advice.

Can I use a free LLM API in production?

No, you shouldn't. Free tiers exist for prototyping; they aren't a stable, supported backend, and several providers explicitly forbid production use. If you build something real on top of free tiers, swap in a paid API before you ship.

Is there a free Claude API?

Not from Anthropic directly. You can get free access to many other strong models (Llama, Qwen, GLM, DeepSeek, GPT-OSS) through the providers in this list. If you want to use Claude itself efficiently, the better move is to keep Claude as the orchestrator and delegate cheap work to a free pool.

How many free LLM tokens can I get per month?

Stacked across providers, the combined free tiers run into the billions of tokens per month, though no single provider gives you that. The practical limit is per-provider rate caps, not a monthly total. Pooling several providers and failing over between them is what makes the combined free budget usable.

Do these free models support tool calling?

Many do. Of the free models in this catalog, about 79 support tool calling and 15 support vision. If your app needs function calls or image input, filter to the capable models first rather than assuming a given free model supports them.

The honest version

Free LLM APIs are one of the best deals in software right now, and also one of the most oversold. The tiers are real, but they're for learning and prototyping, not for running a business on. Providers offer them so you'll build against their models; that goodwill evaporates the second a free tier becomes somebody's production backend.

So use them the way they're meant to be used. Prototype, learn, stack them into a pool, let your expensive model lean on them for the boring work. When you ship something real, pay for the API. This list stays current because the catalog behind it gets re-probed, so check back as providers come and go.

Want the patterns we use to run this in production? Come hang out in the WotAI community with 760+ other builders, and grab the newsletter for the weekly breakdown.