The October 20 release adds session-aware routing, multi-tenant priority/fairness controls, turn-priority scheduling, Responses API state, stable v1 configuration, high-availability routing and coordinated prefill/decode rollouts.
For long agent sessions, context location can affect the next scheduling decision: routing to reusable context may compete with another tenant's claim on the same capacity. Coordinated updates also matter when sessions outlive the workers serving them. These are architectural implications of the announced controls; the release does not quantify update-transition costs or cache loss.
Keep the performance claims separate. One context-heavy demonstration reports 5–10× lower token cost; another, 3× output throughput. A separate peak exceeds 6.6 million output tokens per minute while serving up to 3,000 coding agents with zero preemptions. These are project/vendor demonstrations, not a combined guarantee or independent benchmark. Release and demonstration details.