OpenAI released GPT-6 Astra on September 3, 2026, describing it as its most capable and aligned model to date and, in remarks from president Greg Brockman, a step into what the company called the “AGI era.” The model posted headline scores including 99.9 on ARC-AGI-3 and strong results on computer-use, coding, math, and cybersecurity benchmarks, according to company materials and contemporaneous coverage.
Astra is the first OpenAI system given a “Critical” cybersecurity classification. Because of that rating, the initial rollout went to vetted Daybreak partners—state cyber defenders and operators of critical systems such as water infrastructure—rather than to every paying ChatGPT subscriber at once. OpenAI said broader ChatGPT and API access would follow in the days after launch.
The staggered release proved messy. CEO Sam Altman apologized publicly after paying Plus, Pro, Business, and Enterprise users found themselves locked out of the new flagship. The company began issuing daily usage-quota compensation. Fortune also reported that OpenAI quietly revised evaluation metrics after the announcement in ways that appeared to favor Astra.
Safety documentation accompanying the release said Astra completes more work without surfacing its reasoning, shrinking the oversight surface even as agent sessions get longer and cheaper. That design choice is already shaping the debate over how much of a frontier model’s internal process operators should be able to inspect.
Key takeaway. GPT-6 Astra is both a capability leap and a containment test: OpenAI is shipping a model it rates Critical for cyber risk first to defenders, while paying consumers wait and evaluators watch how little reasoning the system now exposes.
Photo: Unsplash. Sources: OpenAI announcements; Reuters; Fortune; AI Briefing; AIBARS; ThursdAI September 2026 releases.
