# Lanes

> Cheap, fast, OpenAI-compatible inference for open-weight models — priced 2–10× below the managed-API anchors we track, depending on model (blended 3:1 in:out). Agent-native: API keys are first-class and agents are the intended callers.

Point any OpenAI SDK at the base URL below with a `lanes-sk-` key; the request/response shapes are identical to OpenAI's. Two lanes: realtime and batch (runs within 12 hours; per model, batch input is the lower of $0.05/1M and the realtime input rate, batch output is half the realtime output price floored at $0.13/1M and never above realtime — billed on the tokens generated). Realtime first-token latency: the measured p95 per model appears on its line above (nonced measurements, see pricing/basis.json); models without a figure have no measured latency.

- Base URL: https://34.134.52.0.nip.io/v1
- Auth: `Authorization: Bearer lanes-sk-...` (sign in with Google, Microsoft, GitHub or email for a key; free trial credit, no card)
- OpenAPI: https://34.134.52.0.nip.io/openapi.json
- Integrations (which tools work today, and which cannot): https://34.134.52.0.nip.io/integrations

## Models (live)
- `qwen` (Qwen/Qwen3-4B-AWQ) — $0.015 in / $0.200 out per 1M tokens — realtime+batch — batch $0.015 in / $0.130 out per 1M — measured realtime first-token p95 0.515s — 8,192 token context
- `qwen3-8b` (Qwen/Qwen3-8B-AWQ) — $0.030 in / $0.350 out per 1M tokens — realtime+batch — batch $0.030 in / $0.175 out per 1M — measured realtime first-token p95 1.076s — 8,192 token context

More soon: `qwen3-27b`, `qwen3-32b`, `qwen3-coder`, `llama3-70b`, `deepseek-v3`.

## Endpoints
- `POST /v1/chat/completions` — chat completions. Streaming (`stream: true`), structured output (`response_format` json_schema) and tools / `tool_choice` are supported.
- `POST /v1/completions` — text completions.
- `GET /v1/models` — list available models.
- `POST /v1/jobs` — submit a batch job, returns `{"id": ...}`; `GET /v1/jobs/{id}` — poll state/result.

Not supported: `/v1/messages` (Anthropic), `/v1/responses` (OpenAI) and `/v1/embeddings` return 404. Chat-completions clients work as-is; Claude Code and Codex CLI speak only the two unsupported wires and need a translating proxy. Context windows are per model, listed above — a tool that demands more than a SKU offers cannot run on it. Full list: https://34.134.52.0.nip.io/integrations

## Quickstart (Python)
```python
from openai import OpenAI
lanes = OpenAI(base_url="https://34.134.52.0.nip.io/v1", api_key="lanes-sk-...")
r = lanes.chat.completions.create(model="qwen3-8b", messages=[{"role":"user","content":"Hi"}])
print(r.choices[0].message.content)
```