A coding agent makes hundreds of model calls in an afternoon: it reads files, plans, runs a command, reads the output and tries again. On a paid model that adds up before you know whether the agent suits your work. Runcrate now has a free tier for it: five open models you can point Claude Code, Codex, OpenCode, Cline, Kilo Code or Roo Code at, free with your GitHub account.
The models
| Model | Id |
|---|---|
| DeepSeek V4.1 Flash, with its 1017 update | deepseek-ai/DeepSeek-V4.1-Flash, deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| GLM 5.3 | zai-org/GLM-5.3 |
| GLM 5.3 Flash | zai-org/GLM-5.3-Flash |
| gpt-oss 120B | openai/gpt-oss-120b |
| Qwen3.8 Flash | Qwen/Qwen3.8-Flash |
They are the same models, ids and endpoints as the paid API: Chat Completions and Responses at https://api.runcrate.ai/v1 (Codex uses Responses), and Anthropic Messages at https://api.runcrate.ai for Claude Code. GET /v1/models marks each one with free_tier: true, so an agent can find them on its own.
Who gets it
On the API keys page of a workspace you own, choose Sign in with GitHub, then create a key. The GitHub account has to be at least 30 days old, and each GitHub account gets one allowance. The free usage belongs to the workspace: every key and every member of it shares it. A free request costs nothing and does not count toward a key's spend limit.
How free usage is counted
Free usage comes in two parts: a current session, which starts with your first free request, and a week, which resets every Monday at 00:00 UTC. Whichever runs out first stops free requests until it resets.
Every request counts: the prompt, cached tokens included, and the reply. Longer requests and larger models use more of it. A very large prompt is never free: with credit it is billed at the paid price, and without credit the API answers 413 with free_prompt_too_large. Your API keys page shows how much of the session and the week you have used, and when each resets.
Past your free usage
A workspace with credit carries on without a change on your side, at half the list price on shared capacity. When shared capacity is busy and another server takes the request, it is billed at the list price. That price is the same for everyone who pays for these five models, on the free tier or not.
Without credit, requests stop until your usage resets. The API answers 429 with the code free_allowance_used, the time it resets and a Retry-After header, so an agent can wait or tell you to add credit. When free capacity for a model is busy, a workspace with credit is served at the paid price instead; without credit, the API answers free_capacity_busy with a Retry-After header saying when to try again.
Free requests are served by third parties, who handle prompts and outputs under their own terms. The privacy policy says so.
Call it
The free models answer the same requests as any other model. With an API key from that workspace:
curl https://api.runcrate.ai/v1/chat/completions \ -H "Authorization: Bearer $RUNCRATE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek-ai/DeepSeek-V4.1-Flash", "messages": [{"role": "user", "content": "Write a pytest for a function that parses ISO dates."}] }'Any OpenAI SDK works the same way with base_url set to https://api.runcrate.ai/v1.
Ask your agent
Paste this into Claude Code. It reads the steps written for it, asks for your API key and sets itself up on DeepSeek V4.1 Flash; the second part adds the Runcrate MCP server, so the same agent can start GPU jobs:
Set yourself up to use Runcrate's DeepSeek V4.1 Flash: run curl -fsSL "https://runcrate.ai/agents.md?model=deepseek-ai%2FDeepSeek-V4.1-Flash&intent=agent" and follow it. Then add Runcrate's MCP server to yourself: run curl -fsSL "https://runcrate.ai/mcp.md" and follow it.
By hand, Claude Code needs these variables (the Claude Code docs have the full setup):
export ANTHROPIC_BASE_URL=https://api.runcrate.aiexport ANTHROPIC_AUTH_TOKEN="$RUNCRATE_API_KEY"export ANTHROPIC_MODEL=deepseek-ai/DeepSeek-V4.1-Flashexport ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-ai/DeepSeek-V4.1-Flashexport ANTHROPIC_SMALL_FAST_MODEL=deepseek-ai/DeepSeek-V4.1-FlashclaudeWithout the two small-model variables, Claude Code's background calls ask for a model the API does not have. The docs also set CLAUDE_CODE_MAX_CONTEXT_TOKENS to the model's full context; prompts that large are billed, as above. Claude Code also prints its own cost estimate at Anthropic's prices, which is not what you were billed: read your usage on Runcrate instead.
In our Claude Code recipe, Claude Code on DeepSeek V4.1 Flash ran a CUDA test suite on a GPU job through the MCP server. Its 15 model turns cost nothing, and the GPU job $0.03.
See the models and their paid prices on the free inference page, or get an API key and point your agent at it.