Two new Realtime API models cut p95 latency by at least 25%, add configurable reasoning effort, and sharpen noise handling for production voice-agent workflows.
Why it matters: Voice agent infrastructure just got materially cheaper and more reliable — worth revisiting any build that ruled out voice UX on latency grounds.
Today is the last day Fable 5 draws from Pro/Max/Team/Enterprise subscription limits; from tomorrow it bills at $10/$50 per million tokens (double Opus 4.8) and cuts off silently if usage credits aren't pre-enabled.
Why it matters: Any workflow running Fable 5 through a subscription needs usage credits enabled before tomorrow or access stops with no grace period.
OpenAI replaced GPT-5.3 Instant Mini with GPT-5.5 Instant Mini as the model users land on after exhausting their GPT-5.5 Instant or Auto limits — better intent tracking, calibrated tone, and fewer factual errors.
Why it matters: Teams using ChatGPT for content workflows see a quality lift in rate-limited sessions with no plan or prompt changes required.
Version 2.1.202 adds a per-project "Dynamic workflow size" setting (small / medium / large) in /config to dial agent fan-out in workflow runs without touching code.
Why it matters: Developers can now right-size agent fan-out per project; smaller settings cut meaningful token spend on tasks that don't need broad parallel coverage.