GPT-6 Astra: Change Reasoning Effort Without Losing Cache

OpenAI shipped GPT-6 Astra on September 3, 2026. Most of the launch coverage went to two API features: async tool calling and mid-turn steering over WebSockets. The third one got almost no attention, and it is the one that will show up on your invoice.

It is a new input item type called configuration_update. It lets you change how hard Astra thinks partway through a conversation without touching the request-level reasoning.effort parameter. That distinction sounds like trivia. It is the difference between a conversation that reads from prompt cache at $1 per million tokens and one that reprocesses the whole prefix at $10 per million.