The important OpenAI change is not just a new model – it is a more steerable agent runtime
OpenAI’s GPT-6 Astra release matters for more than model quality.
The real architecture shift is in the Responses API: async tool calls, mid-turn steering over WebSockets, and dynamic reasoning effort while preserving prompt cache. That makes long-running agents more concurrent, interruptible, and potentially cheaper per completed outcome.
Astra is expensive at $10/M input and $50/M output tokens, so routing matters. The right benchmark is not tokens per answer – it is total cost per accepted result after retries and human intervention.
OpenAI also added stronger enterprise browser/computer-use policies and named external viewers for ChatGPT Sites. Sites, however, is still unavailable in the EEA, Switzerland, and UK at launch.
Andrea Comba
5 September 2026
