OpenAI ships GPT-6 Astra and calls it the start of AGI

OpenAI shipped GPT-6 Astra today, and president Greg Brockman greeted it with "Welcome to the AGI era" — a line the company has spent years carefully avoiding.

The pitch is computer use: Astra drives browsers, spreadsheets and desktop apps the way a person does, without a bespoke API integration per tool. The numbers are loud — 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 96% on GPQA Diamond, and a clean 100% on ExploitBench. It's OpenAI's largest training run yet, the first pre-trained on 100,000+ GPUs at Stargate, and the first where an earlier model (GPT-5.6 Sol) meaningfully supervised the training.

That ExploitBench score is also the catch. Astra is the first model OpenAI has classified "Critical" for cybersecurity under its Preparedness Framework — it can find unknown vulnerabilities and build exploit chains without a human steering each step. So the rollout is gated: enterprise customers in the Daybreak program first, then Plus, Pro, Business and Enterprise over the coming days, with the sharpest cyber capabilities reserved for vetted defenders. API pricing is $10/M input, $50/M output ($20/$100 in fast mode), which OpenAI claims works out ~57% cheaper per task than Sol.

Why it matters: the first model dangerous enough to ship behind a background check is the same one OpenAI is calling AGI.