Free LLM API keys for testing in 2026

You don't need a credit card to prototype with most major LLM providers. The free tiers are usually generous enough for a weekend, sometimes a month. Here's the honest landscape, and where each provider's "free" actually starts and ends.

Google Gemini

The most generous free tier of any major provider. AI Studio (aistudio.google.com/app/apikey) issues a key tied to a free GCP project with real RPM and TPM caps on Flash. No credit card. The catch: rate limits are aggressive enough that any real chat product will trip them on a slow Tuesday. Still, "free chat completions for prototyping" is a real thing on Gemini.

Groq

Sign up at console.groq.com with a Google account, no card. The free tier gives you per-day token caps per model. Llama 3.3 70B at 300+ tokens/sec for free is a stunning bargain. Limits reset daily, so you can't sustain production traffic, but it's perfect for development.

OpenRouter

OpenRouter exposes a handful of community-hosted models for free (with watermarking), plus low-cost routing across paid models. Useful if you want to bounce between Llama, Mistral, and Gemma without standing up three separate accounts.

Hugging Face

The Inference API is free for hundreds of thousands of models. Rate limits are per-user-per-hour and pretty tight. Cold-start delay is real (the first call can take 10+ seconds while the model loads). Pro accounts get higher caps. Great for occasional inference; not great for any latency-sensitive flow.

Ollama (running locally)

The actual cheapest option: run models on your own machine. Ollama on Apple Silicon runs Llama 3.2 3B and Phi-3-mini comfortably at usable speeds. Zero cost, full privacy, terrible if you don't have the RAM. Test your Ollama daemon to confirm it's reachable.

OpenAI & Anthropic starter credits

Both ship small one-time starter credits to new accounts. They expire after a few months — calendar-time expiry, not on usage. Treat as "play money for the first weekend"; they won't carry a real workload.

What's not actually free

• Replicate, Fireworks, Together, Mistral, DeepSeek, xAI, Perplexity, Cohere — all have promotional credits that come and go, but no permanent free tier you can build on.
• "Free with attribution" model listings on third-party platforms are usually paid behind the scenes.

What to do next

Pick one free option (Gemini Flash or Groq are the pragmatic picks) for development, and budget for paid keys before you ship to real users. Use the comparison table to pick the right paid provider when you're ready.