The cheapest model that still passes.
One endpoint in front of every provider. Every request routed on what it actually needs, not on what you happened to pick six months ago.
Most teams choose one frontier model and send everything to it. Classification, extraction, summarisation, formatting, routine generation, all at frontier prices, when a smaller model returns an answer your users can't tell apart. The router fixes that per request, not per project.
How a request is routed
- 1Request
- 2Classify task and difficulty
- 3Quality bar met by the cheap tier?
- 4Small, fast model or frontier model
- 5Quality score
- 6Response
Every request scored and routed. Quality measured continuously. Below threshold, the route rolls back automatically.
What makes it different
- Quality-scored, not guessed
- Every routed response is scored against the baseline. Scores are visible per feature, per model, per day.
- Automatic rollback
- If quality drops below your threshold, that route reverts and you get an alert. No human in the loop required.
- Per-tenant policy
- Route your enterprise customers to a different model tier than your free tier. Route regulated data to a specific provider or region.
- Hard budget ceilings
- A feature, team or agent can be given a spend cap that the router enforces at request time, not a dashboard you read afterwards.
- Live fallback chains
- Provider outage, rate limit or latency spike, and traffic moves without a deploy.
- Provider arbitrage
- Prices change. The router re-evaluates continuously rather than at contract renewal.
What you keep
OpenAI-compatible API. Change the base URL, keep your code. Streaming, tool use, structured outputs and vision all pass through. Every routing decision is logged with the model chosen, the reason and the price.