# RouterLane

> RouterLane is an LLM API for coding agents where the client picks a tier instead of a model. Send `fast`, `normal`, `smart` or `smart-plus` as the model name; a small decision model reads each new task (200 to 350 ms) and the tier's rules choose the model and reasoning effort. Anthropic Messages, OpenAI Chat Completions and OpenAI Responses formats are supported.

## API essentials

- Base URL for Anthropic clients: https://api.routerlane.com
- Base URL for OpenAI clients: https://api.routerlane.com/v1
- Endpoints: `POST /v1/messages`, `POST /v1/messages/count_tokens`, `POST /v1/chat/completions`, `POST /v1/responses`, `GET /v1/models`
- Auth: an API key starting `rl_`, sent as `x-api-key` or `Authorization: Bearer`. Keys are issued on request at https://routerlane.com/contact/
- Model IDs: `fast`, `normal`, `smart`, `smart-plus` (also `smart+`), optionally prefixed `routerlane/`. Claude-style names map by family: haiku to fast, sonnet to normal, opus to smart.
- Effort: the router replaces any effort or thinking setting the client sends. Anthropic format: thinking off or a budget of 1K, 4K, 12K or 24K tokens. OpenAI formats: the model's own reasoning effort value.
- Response headers on every answered request: `x-router-tier`, `x-router-model`, `x-router-effort`, `x-router-source`, `x-router-request-id`. Cache hits add `x-router-cache: hit`.
- Errors use the format of the endpoint called: 400 bad body or unknown model, 401 bad key, 403 tier not in plan or blocked by a content filter, 429 rate, concurrency or monthly limit, 5xx when every model tried failed.

## Tiers

- Fast: DeepSeek 4.1 Flash, effort off or medium. $0.15 input, $0.60 output, $0.03 cached input per 1M tokens.
- Normal: DeepSeek 4.1 Flash, GLM 5.3 for hard tasks, effort off to high. $0.40, $1.60, $0.08.
- Smart: GLM 5.3, DeepSeek 4.1 Flash for trivial turns, effort low to high. $1.00, $4.00, $0.20.
- Smart+: the client uploads its repository once; setup, explore, build, tests and a QA review run on RouterLane's machine; one patch comes back. Billed at Smart rates. Requires a client with a shell tool. Enterprise plans.

## Docs

- [Overview and quickstart](https://routerlane.com/docs/): base URLs, first requests in curl, Python and TypeScript
- [Authentication](https://routerlane.com/docs/authentication/): keys, headers, plans and tiers
- [Tiers and model IDs](https://routerlane.com/docs/tiers/): accepted names, aliases, effort handling, decision headers
- [API reference](https://routerlane.com/docs/api/): every endpoint with examples, streaming, the /v1/models shape
- [Integrations](https://routerlane.com/docs/integrations/): Claude Code, Codex, pi, OpenAI SDK, Anthropic SDK, Vercel AI SDK, LangChain
- [Smart+](https://routerlane.com/docs/smart-plus/): triggering a job, the upload and patch endpoints, limits
- [Errors and limits](https://routerlane.com/docs/errors/): status codes, error bodies, plan limits, retries

## More

- [How routing works](https://routerlane.com/tiers/)
- [Models and benchmark](https://routerlane.com/models/)
- [Pricing](https://routerlane.com/pricing/)
- [Security and data handling](https://routerlane.com/security/)
- [Status](https://routerlane.com/status/)