Free LLM API keys in 2026: Groq, Gemini, Mistral, OpenRouter and NVIDIA compared

You can still run real work on free LLM API keys in 2026 - but every free tier has a ceiling, and the ceilings are very different. Some limit you per minute, some per day, some per month. Here's what each major provider allows, checked in October 2026, and how to combine them so one limit doesn't stop your team.

Free tiers change without notice. Treat the numbers below as a snapshot, and check each provider's own dashboard before you depend on them.

Groq: the fastest, with the tightest limits

Groq serves open models like GPT-OSS 120B, GPT-OSS 20B and Qwen3.8 27B very quickly. On the free plan each of those models allows 30 requests a minute, 1,000 a day, 8,000 tokens a minute and 200,000 tokens a day, per organization (Groq rate limits).

• Good for: quick answers, chat, short tasks.
• Watch out for: the 8,000 tokens per minute. One long request with a big document or a lot of history can use most of a minute's allowance on its own.

Google Gemini: generous, but the numbers live in your dashboard

Gemini's API has a free tier for its Flash and Flash-Lite text models. Google doesn't print the free limits in its documentation - they're shown per project in Google AI Studio. Third-party trackers put Flash-Lite at roughly a thousand or more requests a day, with much lower daily limits on the stronger Flash models.

• Good for: everyday work, long documents (large context windows).
• Watch out for: image generation models are paid-only, and limits can drop without warning.

OpenRouter: one key, many free models

OpenRouter gives access to many models, and the ones with ":free" in their ID cost nothing. Free variants allow 20 requests a minute and 50 a day; after a one-time purchase of $10 in credits that rises to 1,000 a day (OpenRouter limits).

• Good for: trying different models with one key.
• Watch out for: 50 a day goes fast for a team.

Mistral: the biggest monthly allowance

Mistral's free Experiment plan needs a phone number but no card, and gives rate-limited access to its models with a reported allowance of around a billion tokens a month at roughly one request a second. Check your exact limits in Mistral's console.

• Good for: volume - long tasks, background work, heavy users.
• Watch out for: the free plan may use your data to improve Mistral's models. Turn that off in your Mistral privacy settings before using it for work.

NVIDIA Build: many models, a per-minute cap

NVIDIA's build.nvidia.com hosts a large catalogue of open models behind one free API key. Free accounts are limited to about 40 requests a minute, with NVIDIA's moderators saying there's no official way to raise it on the free tier.

• Good for: steady throughput across many models.

Images: Cloudflare Workers AI

For image generation, Cloudflare's Workers AI includes 10,000 "neurons" a day for free (Workers AI pricing). With a fast model like FLUX.2 [klein] at 1024x1024, that works out to roughly 95 images a day per account.

What's no longer free

• GitHub Models was retired on 30 July 2026.
• Cerebras now asks for a card.

How to combine free keys (honestly)

No single free tier is enough for a busy team, but together they go a long way:

• One account per person. Limits are per account, so each teammate adding their own free keys multiplies what the team can do. Creating several accounts for yourself is against every provider's terms - don't.
• Use the small model for quick asks. Free tiers give far more of the small models than the strong ones. Save the strong model for real work.
• Move on when a key runs out. A request refused for a rate limit should go to the next key, not fail - and the refused key should rest for as long as the provider asked.

How Teamski uses them

In Teamski, each person can add their own free keys, then pick Rotate my keys (free first) as a channel's model. Teamski uses the free tiers first, picks the small or strong model for each message, skips a key when a request is too big for its free tier, and moves to the next key when one hits its limit. The team works in shared channels with one AI agent each - see what a shared LLM workspace is.