OpenAI’s new flagship model scores 99.9% on ARC-AGI-3 and 98% on FrontierMath Tier 4, prompting President Greg Brockman to declare the AGI era officially begun — but independent evaluations reveal a more nuanced picture.
OpenAI released GPT-6 Astra on September 3, 2026, calling it “the world’s most intelligent and best-aligned model” . OpenAI President Greg Brockman told reporters that “it is not unreasonable to say we are now in the AGI era,” adding that future historians may pinpoint Astra as the moment AGI arrived .
The model’s self-reported benchmark scores are staggering: 99.9% on ARC-AGI-3 (up from 7.8% for GPT-5.6 Sol), 98% on FrontierMath Tier 4, and 100% on ExploitBench . In mathematics, Astra contributed to a paper proving tighter bounds on prime gaps, reducing the upper limit from 246 to 186 — a result formalized in Lean .
However, independent testing organization Artificial Analysis found Astra Max scoring 61 on its Intelligence Index, tied with GPT-5.6 Sol but below Anthropic’s Claude Fable 5.1 at 66 . The ARC-AGI-3 score also dropped to 62.7% under standard testing frameworks .
Astra’s core differentiator is computer use — the ability to operate software like a human, filling forms, navigating spreadsheets, and managing multi-step workflows without custom APIs . OpenAI reports that in OSWorld 2.0 simulations, Astra completes tasks in approximately 40 minutes compared to Sol’s 75 minutes, with higher accuracy .
