Free LLM API Options in 2026: Limits Checked
Which LLM APIs are still free? We checked OpenRouter, Gemini, Groq, Mistral, Cloudflare, SambaNova and Cerebras docs: real limits, cards, gotchas.

Free LLM API Options in 2026: Every Free Tier, Checked Against Official Docs
Looking for a free LLM API you can actually build on? If you were using MiniMax M3.1 Flash for free on OpenRouter last week, you have already hit the first problem: its free launch period ended on October 22, 2026. It is still on OpenRouter, but now at $0.15 per million input tokens and $0.60 per million output tokens.
So which free AI APIs still work, and what are the real limits? We read the official docs of each provider on October 26, 2026. Below is what we could confirm, what we could not, and the catches that come with almost every free tier.
TL;DR
- Free, no-card options still exist: OpenRouter's
:freemodels, SambaNova Cloud's Free Tier, Google Gemini's Free tier, Mistral's Free mode, and Cloudflare Workers AI's daily free allocation. - Daily caps are low. SambaNova's Free Tier allows 20 requests per day per model, and OpenRouter allows 50 free-model requests per day until you have bought 10 credits. One agent run can use that up.
- GitHub Models was fully retired on July 30, 2026. Cerebras offers a $5, 30-day trial that needs a payment method, not a permanent free tier.
Free LLM API comparison table
| Provider | Free allowance | Rate limit | Card required | Notes |
|---|---|---|---|---|
OpenRouter :free models |
50 requests/day (under 10 credits purchased, all time); 1,000 requests/day (10+ credits purchased) | 20 RPM | No for 50/day; 1,000/day needs 10 credits bought once | OpenAI-style API. Vision and tools depend on the model |
| OpenRouter MiniMax M3.1 Flash | Free period ended Oct 22, 2026 | — | — | Now $0.15/M input, $0.60/M output |
| SambaNova Cloud | Free Tier when no payment method is linked | 20 RPM / 20 RPD / 200K tokens per day, per model | No | Listed for models such as DeepSeek-V3.1, Llama 3.3 70B, gpt-oss-120b |
| Google Gemini API | Free tier on selected models (e.g. Gemini 3.8 Flash, 3.5 Flash-Lite) | Per project; shown in AI Studio, not published as fixed numbers. Daily quota resets at midnight Pacific | No billing account needed for the Free tier | Free-tier content is used to improve Google's products |
| Mistral | Free mode, enabled by default | Usage and rate limits apply; check your console | No | Exact limits not published on the page we checked |
| Cloudflare Workers AI | 10,000 Neurons/day, resets 00:00 UTC | Check dashboard | No (Workers Free plan) | Vision model llama-3.2-11b-vision available. GLM-5.3, GLM-5.3-Flash, DeepSeek V4.1 Flash and some Kimi models need a paid billing method |
| Groq | Free access available (check console) | Published table is the Developer plan's base limits; your free limits are on your Limits page | Check console | Limits apply per organization |
| Cerebras | No permanent free tier: $5 trial credit, expires after 30 days | Trial: 5 RPM, 1M tokens/day per model | Yes (verified payment method) | qwen-3.8-27b accepts images |
| GitHub Models | Retired July 30, 2026 | — | — | No longer available |
Last verified: October 26, 2026, against each provider's official docs (sources at the end). "Check dashboard" or "check console" means the provider does not publish a fixed public number, or we could not confirm one. Free tiers change often, so confirm before you build on one.
What happened to MiniMax M3.1 Flash free?
MiniMax M3.1 Flash was released on October 19, 2026 and got a lot of attention: it is multimodal (text, image and video in), has a 1M-token context window, and was free on OpenRouter during its launch. That free period lasted about three days. Since October 22, OpenRouter serves it at list price only.
You may still see places claiming it is free elsewhere. We could not verify any of those quotas from official sources, so treat MiniMax M3.1 Flash as a paid model now. The good news is that at its current price it is one of the cheaper capable models around.
Free LLM API providers, one by one
OpenRouter: the easiest free OpenAI-compatible API
OpenRouter is the simplest starting point for a free OpenAI-compatible API: one OpenAI-style endpoint, many models, and a rotating set marked :free. Without buying anything you get 50 free-model requests per day. Buy 10 credits once and that rises to 1,000 per day. Free models are capped at 20 requests per minute either way. Which models are free changes often, and MiniMax M3.1 Flash shows how quickly one can disappear.
SambaNova Cloud
SambaNova applies its Free Tier when no payment method is linked. Limits are strict: 20 requests per minute, 20 requests per day, and 200,000 tokens per day, for each model. Good for trying a model, not for anything continuous.
Google Gemini API free tier
Gemini's Free tier needs no billing account and covers selected models, including Gemini 3.8 Flash and Gemini 3.5 Flash-Lite. Newer flagship models such as Gemini 4 Argon are not on the Free tier. Google does not publish fixed free-tier numbers in the docs; you see your limits in AI Studio. Limits are per project, not per API key, and the daily quota resets at midnight Pacific time. One thing to know: on the Free tier, Google says your content is used to improve its products.
Mistral
Mistral's Studio starts you in Free mode with API access enabled and no credit card. Mistral says usage and rate limits apply, but the page we checked does not list the numbers, so check your console.
Cloudflare Workers AI
The Workers Free plan includes 10,000 Neurons per day, reset at 00:00 UTC. Neurons are Cloudflare's own usage unit, so how many requests that covers depends on the model. Some of the newest models, including GLM-5.3, GLM-5.3-Flash and DeepSeek V4.1 Flash, require a paid billing method.
Groq
Groq is known for very fast inference. Its rate-limit page lists per-model numbers, but says they are the base limits for the Developer plan, so the free limits on your account may differ. Check the Limits page in your Groq console.
Cerebras and GitHub Models
Cerebras gives new accounts $5 in credits after you add a verified payment method. The credits expire 30 days after they are granted, and Cerebras says it has no permanent free tier. GitHub Models, which many people used as a free GPT API, was fully retired on July 30, 2026.
The catch: limits of free AI APIs
Free tiers are great for experiments. Before you wire one into a product or an agent, know the common catches:
- Free can end with little notice. MiniMax M3.1 Flash was free on OpenRouter for about three days. GitHub Models was retired entirely. Have a fallback.
- Daily request caps are tiny for agents. An agent that plans, calls tools and retries can make 20 to 50 calls on one task. That is a full day on SambaNova's Free Tier, or on OpenRouter without credits.
- "Free" sometimes means "trial." Cerebras needs a payment method and the credit expires. Read the terms first.
- New, popular models are often excluded. On Cloudflare, GLM-5.3 and DeepSeek V4.1 Flash need a paid billing method. On Gemini, the newest flagship is not on the Free tier.
- More keys do not mean more quota. OpenRouter says extra accounts or API keys will not change your rate limits. Gemini counts limits per project, not per key. Groq and Cerebras count them per organization.
- Data use terms differ. Google says Gemini Free-tier content is used to improve its products. Read each provider's terms before sending anything sensitive.
- Every provider is a separate integration. Different keys, base URLs, model names and error formats. Switching when one runs dry costs engineering time.
How to get a free LLM API key safely
Getting a free LLM API key is usually a few minutes' signup. A few habits save trouble later: keep keys in environment variables, not in code; set an expiry and rotate keys where the provider allows it; set a spending limit if the provider asks for a card; and log which provider served each request, so you can tell a rate-limit error from a real bug.
When free stops being enough
Once your agent needs more than a few dozen calls a day, free tiers start costing you time instead of money. That is the problem XCoffee is built for.
XCoffee plan details updated November 12, 2026.
XCoffee is a model API relay for agent work:
- One key, several API formats. Clients using Chat Completions, Responses, Anthropic Messages or Gemini native generation can use the matching endpoint with the same key. Set the base URL to
https://api.xcoff.ee/v1and use a group name as the model. See the setup guide. Running Claude Code skills like tstack? See how to run tstack on one cheap API key. - Groups by task, not by model. Choose Coding (
xcoffee-coding), Fast (xcoffee-fast) or Default (xcoffee). The live model-group list shows current credit multipliers and some member models. - Automatic failover. If an upstream hits an eligible rate limit or fails before output begins, the request is tried on another available upstream in the same group.
- Plans for different workloads. A 30-day pass includes a daily credit allowance and a rolling hourly cap. Prices, allowances and payment options depend on the plan; see the current plans. Daily credits reset at UTC 00:00.
- Usage depends on the request mix. Credits account for uncached input, cache reads and output, including reported reasoning, at the selected group’s multiplier. Cache writes are charged as full input. A credit allowance is not a fixed number of total tokens; check the current credit formula.
What to know before you connect, straight from our homepage FAQ:
- Concurrency and request-rate limits depend on your plan and are shared across your account’s keys, groups and protocols.
- Each plan has its own rolling hourly credit and request allowances. These recover as older usage leaves the window. Your account shows the limits that apply to your pass.
- Unused daily credits do not roll over.
- Client tools are supported; stored sessions, uploaded file APIs and provider-hosted tools are not included.
- You pick a group, not an individual member model.
If MiniMax M3.1 Flash was your favourite free model, this is a direct way to keep using it at a flat, predictable price.
FAQ
Is there a completely free LLM API in 2026?
Yes. OpenRouter's :free models, SambaNova Cloud's Free Tier, Gemini's Free tier, Mistral's Free mode and Cloudflare Workers AI's daily allocation all work without paying. All of them have tight daily or per-minute limits.
Is MiniMax M3.1 Flash still free?
Not on OpenRouter. Its free launch period ended on October 22, 2026, and it now costs $0.15 per million input tokens and $0.60 per million output tokens there. We could not verify free access anywhere else from official sources.
Which free LLM API is OpenAI-compatible?
OpenRouter is the clearest choice: its API follows the OpenAI format and includes free models. For other providers, check their docs for an OpenAI-compatible endpoint.
Do I need a credit card for a free AI API key?
Usually not. OpenRouter, SambaNova, Gemini, Mistral and Cloudflare's Workers Free plan do not need one. Cerebras does, for its trial credit.
Is there still a free GPT API?
GitHub Models, a common free route to GPT models, was retired on July 30, 2026. Today the practical free options are open-weight models on OpenRouter, SambaNova or Cloudflare, or Gemini's Free tier.
Does creating more API keys increase my free quota?
No. OpenRouter says extra accounts or keys do not change your rate limits, Gemini limits are per project rather than per key, and Groq and Cerebras apply limits per organization.
Which free API supports image input?
Gemini's Free-tier models accept images. Cloudflare Workers AI has vision models such as llama-3.2-11b-vision. On OpenRouter it depends on the specific free model.
What is the cheapest way to run an AI agent after free limits run out?
If you prefer a predictable subscription budget, compare XCoffee’s current plans. One key supports the matching OpenAI, Anthropic and Gemini endpoints, with task-based groups and retry on eligible errors before generation. Daily allowances and hourly caps vary by plan.
How we verify
We check each provider's official docs and update this page when limits change. If a number is not published or we could not confirm it, we say "check dashboard" instead of guessing. Sources, checked October 26, 2026:
- OpenRouter limits: https://openrouter.ai/docs/api-reference/limits
- OpenRouter MiniMax M3.1 Flash: https://openrouter.ai/minimax/minimax-m3.1-flash:free
- SambaNova rate limits: https://docs.sambanova.ai/docs/en/models/rate-limits
- Gemini rate limits: https://ai.google.dev/gemini-api/docs/rate-limits
- Gemini pricing: https://ai.google.dev/gemini-api/docs/pricing
- Mistral Studio quickstart: https://docs.mistral.ai/getting-started/quickstarts/studio/activate-and-generate-api-key
- Cloudflare Workers AI pricing: https://developers.cloudflare.com/workers-ai/platform/pricing/
- Groq rate limits: https://console.groq.com/docs/rate-limits
- Cerebras rate limits: https://inference-docs.cerebras.ai/support/rate-limits
- GitHub Models: https://docs.github.com/en/github-models