OpenAI's GPT-6 Astra prompt guide: strong model means you should trim your SKILL and AGENTS.md rules

xiaohu · x · 2026-09-06

OpenAI updated its official prompting guide alongside GPT-6 Astra. Key takeaways from the breakdown:

Model behavior: Astra is positioned as the flagship for computer use, browsing, and software engineering. It handles long multi-step tasks better and is more willing to ask clarifying questions at critical junctures. Cost should be compared per completed task — fewer output tokens may offset higher per-token pricing, but verify on your own workload.
New API capability: async tool calls (async: true) let the model keep reasoning and calling other tools while your app executes a tool; results return via the original callid. Requires app-side integration.
Practical tip: Astra is more sensitive to instructions in context — steps in SKILL.md and conventions in AGENTS.md can affect execution. With a stronger model you may need to substantially trim those rule files. If it keeps stopping despite being told to continue, check the skill and project files it reads.
Migration: check interfaces, reasoning effort, deprecated params, and cache config when upgrading.

More from coding & agent