OpenAI released its GPT-6 Astra model on September 3, 2026. CEO Sam Altman announced the launch via a tweet that received 4.1 million views, listing performance metrics on benchmarks like FrontierMath, ARC-AGI-3, and ExploitBench without using the term AGI.
The phrase 'Welcome to the AGI era' was spoken by OpenAI President Greg Brockman during a private press briefing, not in the official company announcement. Following the event, users reported access issues, leading Altman to apologize for the rollout.
Technical evaluations show Astra uses 'recurrent depth' architecture. On the ARC-AGI-3 benchmark, the model achieved 99.9% using a 'Provider Adapter' to maintain hidden reasoning states, while independent tests by the ARC Prize Foundation on standard settings yielded 62.7%.
Astra is classified as 'Critical' in cybersecurity, having identified two previously unknown vulnerabilities in the V8 engine during internal testing. In safety simulations, the model triggered fewer consistency violations compared to its predecessor, Sol.
Despite high performance in agentic tasks, Astra showed a regression on the GDPval benchmark, which measures economically valuable work. Artificial Analysis reported Astra scored 61 points, trailing behind Anthropic’s Claude Fable 5.1, which leads the index with 66 points.
The definition of AGI remains contested. OpenAI’s original charter defined it as systems outperforming humans in economically valuable work, but recent internal shifts have led to the term being described by leadership as a 'spiritual concept' rather than a technical metric.
