You can build a real product on large language models in 2026 without paying for tokens. Dozens of providers give away API access to strong open models — DeepSeek, Qwen, GLM, Kimi, Nemotron, Gemma — and Google still offers a free tier for Gemini Flash. The catch is that every provider plays by different rules: some cap you at 50 requests a day, some need a credit card, some log your prompts, and free routes appear and disappear every week.
This guide explains how free LLM APIs actually work, which providers are worth your time, how to pick a model for your task, and how to stack several free tiers so you rarely hit a limit. The numbers come from our live catalog, which polls 15 providers every hour.
Three kinds of "free"
When a provider says an LLM API is free, it means one of three things:
- Zero-price routes. The model is listed with an input and output price of $0. OpenRouter marks them with a
:freesuffix; UnoRouter, Nous Portal, TokenRouter, and Vercel AI Gateway do the same. You pay nothing per token, but there is a request limit. - Developer free tiers. The provider gives a standing monthly or daily allowance: Google AI Studio for Gemini, Groq, Cloudflare Workers AI (10,000 Neurons per day), NVIDIA Build for prototyping, Cohere trial keys (1,000 calls per month).
- Trial credits. A one-time balance for new accounts — Cerebras, AI21, Fireworks, and the big clouds. Useful for experiments, not for a product. We track these separately on the AI credits page.
The providers worth knowing
| Provider | What is free | Card | Best for |
|---|---|---|---|
| OpenRouter | :free models: 50 requests/day, or 1,000/day after a one-time $10 top-up; 20 requests/min | No | One key for many models |
| UnoRouter | The largest catalog of :free routes, shared capacity and per-model limits | No | Trying many new models |
| NVIDIA Build | Free serverless inference for development | No | Fresh open models (GLM, Kimi, DeepSeek, Nemotron) |
| Google AI Studio | Gemini Flash and Flash-Lite with rate limits | No | Multimodal input, long context |
| Groq | Per-model request and token limits per minute and per day | No | Very fast responses |
| Cerebras | Free trial with daily token limits | Yes | Speed |
| Cloudflare Workers AI | 10,000 Neurons per day | No | Edge apps, small models |
| Nous Portal | :free models on the $0 plan | No | Agent workloads (Hermes Agent) |
| Vercel AI Gateway | Zero-priced promotional models | Yes | Apps already on Vercel |
The full comparison, including base URLs for every provider, is on the providers page.
Which free models are strongest right now
We match every free model with the public Artificial Analysis leaderboard. Its Intelligence Index combines ten benchmarks covering reasoning, knowledge, coding, and agentic tasks. The top free models today:
| Model | AA Intelligence Index | Context | Free at |
|---|---|---|---|
| GLM 5.3 | 44.8 | 1M | NVIDIA Build, UnoRouter |
| Kimi K3 | 43.6 | 1M | NVIDIA Build, UnoRouter |
| GLM 5.3 Flash | 41.8 | 1M | NVIDIA Build, OrcaRouter, UnoRouter |
| Qwen 3.8 Flash Next | 39.8 | 1M | UnoRouter |
| DeepSeek V4.1 Flash | 39.5 | — | NVIDIA Build |
| DeepSeek V4 Pro | 36.0 | 1M | UnoRouter |
Two things stand out. First, the best free models are open-weight Chinese MoE models with million-token context windows. Second, the same model is often free at several providers, each with its own limits — which is exactly what makes stacking possible.
Pick a model by task
- Coding and IDE agents: GLM 5.3, Kimi K3, DeepSeek V4, Qwen 3.8. See free coding models and our coding ranking.
- Autonomous agents: models with reliable tool calling — free models for agents.
- Large documents and repositories: long-context models with 256K to 1M tokens.
- Screenshots, OCR, images as input: vision models such as Gemini Flash, Gemma 4, and Kimi K3.
- Low latency: small models on Groq or Cerebras.
Your first request
Almost every provider speaks the OpenAI Chat Completions protocol, so one snippet works everywhere — only the base URL, key, and model ID change. Here is OpenRouter:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "z-ai/glm-5.2:free",
"messages": [{"role": "user", "content": "Explain MoE in two sentences."}]
}'Every model card in the catalog has a ready-to-copy snippet in cURL, Python, and JavaScript for the best available provider.
The catches
- Rate limits. Free routes run on shared capacity. Expect HTTP 429 errors at peak times and design for retries.
- Privacy. Some free routes may use your prompts to improve models. Google says free-tier Gemini content may be used to improve its products. Do not send customer data through free endpoints without reading the provider's policy.
- Churn. Promotional models vanish without notice. Our new models feed and RSS show what appeared and what disappeared in the last weeks.
- Hidden card requirements. Cerebras and Vercel AI Gateway ask for a payment card even for zero-price models.
- Context caps. A router may serve a 1M-context model with a smaller window. The catalog shows the provider-specific value.
A free stack that rarely runs dry
- Create keys at OpenRouter, UnoRouter, NVIDIA Build, and Google AI Studio — none of them needs a card.
- Put them behind one OpenAI-compatible client: most tools (Cline, OpenCode, Open WebUI, LiteLLM) let you define several providers and switch on errors.
- Use a strong model (GLM 5.3 or Kimi K3) for planning and a fast one (Gemini Flash or a Groq-hosted model) for routine calls.
- For privacy-sensitive work, run a model locally with Ollama or LM Studio.
If you only need occasional heavy lifting, compare these free tiers with paid plans in our subscription ROI table — sometimes a $10 plan covers what a week of juggling free keys would.
Keep up with changes
The catalog refreshes every hour, and each model family has its own page with providers, limits, and benchmarks. For tools that plug into these APIs, see free AI software; for one-time credits and startup programs, see AI credits.