Canary rollouts: upgrade models in production without downtime

Together AI's canary rollouts let operators expose a replacement model to a small share of requests before expanding its traffic. The platform checks readiness before routing requests, then can compare latency or error metrics after each stage. A failed metric check pauses the rollout at its current traffic split; cancellation freezes that split, and restoring the previous model requires a reverse rollout. In the company's documented demonstration, a 137% increase in 95th-percentile latency triggered a pause at 10% traffic. Together reports 6,800 requests without a non-200 response through the switch and recovery. The example shows controlled exposure to a slower model, with the recovery decision remaining with the operator.