Flow adaptive routing
Train an adaptive router from your project's model choices.
Availability
Flow is available on all paid plans at 1 Credit per successful training run. Enable it separately for each routing configuration.
Collect and train
Flow and TokenOps use the same project collection settings. Enable collection and send requests with explicit model choices. Auto-selected requests are excluded from Flow training.
Training unlocks when enough eligible examples are available across multiple models. A manual training run uses recent retained requests within the service limits. Your existing version keeps serving while a new one trains.
Review and activate
Review accuracy, macro F1 and the confusion matrix. These describe agreement with historical model choices, not answer quality or guaranteed savings. Rejected versions cannot be activated.
Activate a version, then enable Use Flow in a routing configuration's Cost and quality step. Requests continue to use model auto and the existing routing_config field. No Chat changes or new model identifier are needed.
Activating a replacement updates all configurations using Flow after confirmation. Model and price restrictions still apply. Flow does not use the cost-quality slider; its saved value applies to standard Auto fallback.
Credits and controls
Training reserves 1 Credit from your subscription balance. It is charged only when a usable version is produced; rejected, failed or timed-out runs release the reservation. Activation costs no additional Credits. Serving uses normal inference rates.
Disable Flow for one configuration or the whole project to return to standard Auto. Configuration choices are preserved when pausing the project. Resuming after subscription expiry requires confirmation.
You can restore a previous compatible version or delete inactive versions. Deleting all Flow history disables Flow and preserves captured requests. Clearing captured requests does not delete trained versions.