Docs/Models & routing

Smart routing

How the gateway picks and ranks providers for each request.

On this page

The default policy

On each request the gateway collects every candidate provider for the model, then filters and ranks them:

  1. Availability filter — skip providers that are cooling down, at capacity or failing repeatedly.
  2. Capability filter — skip providers that can't do what this request needs (an image attachment on a text-only provider, say).
  3. Ranking — order by the provider priority the platform configured for the model, then nudge by recent performance: providers doing worse move down.
  4. Call the top candidate, and on failure continue per the failover rules.

Keeping a conversation on one provider

For some models, one conversation stays on the same provider for a while to avoid bouncing between them. If that provider degrades, the gateway moves away at once — consistency never wins over availability.

What you need to do

Nothing. Provider choice, ranking and switching are maintained by the platform; requests don't need to — and can't — name a provider. You only pick the model.

Tip

If a model is consistently slow or flaky for you, contact us with a few request IDs (the X-Maxx-Request-ID response header) and we'll tune that model's provider setup.

Contact us

Something unclear, or need a human? Pick the channel that fits your question.