7 free AI APIs to ship your first AI feature

Your first AI feature does not need a credit card. The work is picking the tier that fits your workload before its rate limits pick for you.

How we picked: each has a standing free tier confirmed in vendor docs this month and an install command from its own docs. We dropped Cerebras (expiring trial) and GitHub Models (retired July 2026).

The 7 free AI APIs at a glance #

1. Gemini API: the free multimodal default #

The Gemini API gives free rate-limited access to Gemini Flash models with text, image, audio and video input (details).

What: one key from Google AI Studio gets free calls to Flash models plus 500 search-grounded requests a day on Gemini 2.5 Flash and Flash-Lite. Why: Google's own models at $0, with image, audio and video input that most free tiers lack. Use when: you are starting a prototype and want the fewest compromises. Install:

pip install google-genai

Catch: free-tier prompts may be reviewed by humans and used to train Google products — keep confidential data on the paid tier.

2. Groq: open weights, no card #

Groq serves open-weight models like gpt-oss-120b on its own inference chips (LPUs) (details).

What: an OpenAI-compatible endpoint at 30 requests/minute and 1,000 requests/day per model, plus free Whisper speech-to-text. Why: 1,000 requests a day is enough to build against all week. Use when: your prototype calls open-weight models at a predictable daily volume. Install:

pip install groq

Catch: limits apply per organization and are tight; the paid Developer plan buys headroom.

3. OpenRouter: one key, many models #

OpenRouter fronts many models behind one OpenAI-compatible key, including variants that cost $0 (details).

What: :free model variants at 20 requests/minute and 50 requests/day on a fresh account — 1,000/day once you have ever bought 10 credits. Why: you swap models in a config line instead of juggling five vendor consoles. Use when: you are comparing models or want a fallback chain. Install:

pip install openai   # then set base_url https://openrouter.ai/api/v1

Catch: 50 requests a day runs out fast without the $10 top-up.

4. Cloudflare Workers AI: free at the edge #

Cloudflare Workers AI runs open models like Llama on Cloudflare's network, close to your users (details).

What: 10,000 free Neurons a day (about 48,000 output tokens on Llama 3.3 70B), callable from a Worker or plain REST. Why: the AI call lives in the same deploy that serves your app. Use when: you already deploy on Workers. Install:

curl https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/meta/llama-3.1-8b-instruct \
  -H 'Authorization: Bearer {API_TOKEN}' \
  -d '{ "prompt": "Where did the phrase Hello World come from" }'

Catch: on the free Workers plan usage stops at the daily allocation, and the largest models (Kimi, GLM, DeepSeek) need the paid plan.

5. Cohere: embeddings and rerank for RAG #

Cohere gives you a free trial key at signup (details).

What: 1,000 free API calls a month across Command chat models, Embed v5 at 2,000 inputs/minute, and Rerank. Why: it is the only free key here with embeddings and reranking, the two parts that decide retrieval quality. Use when: you are building search or RAG (answers grounded in your own documents). Install:

pip install cohere

Catch: trial keys are rate-limited and not permitted for production or commercial use.

6. Mistral: EU-hosted open weights #

Mistral's Free plan includes rate-limited API access in Studio Free mode (details).

What: free calls to Mistral Large 4, Mistral Small 4 (Apache 2.0) and Codestral, no card required. Why: European hosting and open weights with a training-data opt-out, when data residency matters. Use when: your users or your rules push you toward the EU. Install:

pip install mistralai

Catch: exact rate limits are not published — you see them only in your admin console.

7. Free-LLM: the directory for what's next #

nejib1/Free-LLM is an MIT-licensed catalog synced automatically from free-llm.com.

What: lists 86 providers with permanent free tiers, trial credits or no-card signup (950+ stars, November 2026). Why: vendors change terms monthly, and this list tracks them. Use when: you outgrow these seven. Install:

git clone https://github.com/nejib1/Free-LLM

Catch: a catalog, not an API. Check each vendor's terms before you build on it.

Which one to try first #

Starting a chat or multimodal prototype → #1. Latency-sensitive demo → #2. Comparing models before committing → #3. Already deploying on Workers → #4. Building search or RAG → #5. Need EU hosting → #6. Need something these don't cover → #7.