Canary / A-B rollouts

A rollout splits a logical model's traffic across two providers by weight — for example, send 90% of coding-default requests to your current provider and 10% to a new one, then ramp to 50/50 and 100% as you gain confidence. The split is per request (randomized in proportion to the weights), so you can A/B test or safely roll out a new provider or model.

Who can do this

Org admins (for their organization) and platform admins, on Projects → Rollouts.

How it works

The gateway routes the logical model to both providers using weighted routing — each incoming request independently goes to provider A or B in proportion to the weights. There's no stickiness: the same user may hit either provider on successive requests (good for an even A/B test). Removing the rollout returns the model to its single provider immediately.

Configure a rollout

1. Open Projects → Rollouts and pick a logical model (one already configured under Routing).
2. Set Target A and Target B: each is a registered provider, that provider's upstream model id, and a weight. The two weights must sum to 100.
3. Save rollout. Traffic begins splitting within ~1 second.
4. Ramp by editing the weights (e.g. 90/10 → 50/50 → keep the winner). Remove canary returns the model to its single Routing entry.

Tips

• Both targets must be providers you've registered for the project (see Providers).
• Use it to trial a new provider (e.g. a self-hosted model vs a cloud API) or compare two model versions on the same logical name your apps already call — no client changes needed.
• For a gradual per-user rollout (a user consistently gets the new model), ask us — that sticky mode is a separate option.

Next steps

• Routing — the logical models and providers a rollout splits between.
• Providers — register the providers used as rollout targets.