OpenAI AI Updates: September 6, 2026

1. OpenAI Documents How to Stop GPT-6 Astra From Stalling on Clarifying Questions

OpenAI. The GPT-6 Astra model documentation names where the model behaves unexpectedly and supplies prompt text to correct it. Astra asks clarifying questions more often than GPT-5.6 Sol, which OpenAI frames as making it a “more effective collaborator” at the cost of stopping where users expect it to continue. The recommended fix instructs the model to infer intent from context, treat phrases like “can you…”, “I want to…”, and “help me…” as calls to act, and work independently toward the goal unless an action is destructive or irreversible, holding approval requests until it has a concrete reviewable result. The docs also warn that unclear or contradictory instructions in skill files such as AGENTS.md can make Astra block work or veer off, and suggest a debugging prompt that forces it to name and link the exact SKILL.md instruction that caused it to pause. Two other documented quirks: the model delegates to parallel sub-agents less often than expected, and it runs test suites out of proportion to small code changes unless told to rerun tests only when new failures justify it. Source

2. The Astra Prompting Guide Ships a Blocklist of “Slop Words”

OpenAI. Alongside the behavioral guidance, OpenAI publishes a list of phrasing to suppress, which it labels “slop words.” The blocklist covers “Conclusion:” in conclusions, “delve into”, “it’s worth noting”, “what’s important is”, “really/truly”, summary closers such as “In short:…” and “The simplest mental model is:…”, and contrastive constructions like “X, not Y” on the grounds that they introduce an alternative the user never asked for. It also targets invented hyphenated compounds, citing “exact-head checks” and “editorial-row layouts” as examples, and tells the model to state the intended action directly rather than listing what it will not do or how it will categorize results. For technical writing the guidance is to prefer plain language and reference technical detail only where it illustrates the work. Developers migrating existing projects can run the recommended changes through Codex with the OpenAI Docs skill. Source

3. GPT-6 Astra Reaches Pro, Enterprise, and Business Premium at Half the Message Rate

OpenAI. Astra is now available to ChatGPT Pro, Enterprise, and Business Premium subscribers through ChatGPT Work and Codex, with Plus and Business users following in the coming days, and it is reachable through the OpenAI API, Microsoft Azure, and AWS Bedrock. The standard model carries roughly half the messages per five-hour window as GPT-5.6 Sol across every plan: Plus users get an estimated 5 to 45 messages against Sol’s 10 to 100, and the Pro 20x tier gets 100 to 900 against Sol’s 200 to 2,000. GPT-6 Astra Pro allowances count against existing quotas, at 200 messages per week on the $200 Pro plan, 50 per week on the $100 Pro plan and Business Premium, and 15 per month on Business Standard. Free and Go tiers get neither frontier model. Source