Every release, on one axis.

Agents that run longer than a minute now keep their worker warm between steps. Tool calls no longer pay a cold start.

Tool output streams back to the model as it arrives, so a slow search no longer blocks the next token.

Every run can be replayed step by step, with the exact inputs, timings and tool responses it saw.

Keys on Scale and Dedicated plans can mark runs as interactive. Interactive runs jump the batch queue.

Two new regions. Requests are routed to the closest healthy region unless a key pins them.