On one benchmark, OpenAI's previous model scored 7.8. This one scored 99.9
GPT-6 Astra shipped on Thursday. The pitch isn't how well it answers, it's that it operates your computer: opening browsers, moving through spreadsheets, filling forms, carrying a multi-step job across several apps until there's a finished document at the end. Greg Brockman closed the session with "Welcome to the AGI era."
On Tuesday I wrote here about OpenAI quietly buying tens of thousands of Mac minis to train computer-use agents. Whatever those machines were learning, it shipped on Thursday.
The 7.8 against 99.9 is ARC-AGI-3, which drops a model into an environment it has never seen and makes it work out the rules unaided. Claude Opus 5 managed 30.2% on the same test. A jump that size in one generation is not something I've seen in a while.
But there was a second half to Thursday that almost nobody quoted.
On ExploitBench, can it turn a known software vulnerability into a working exploit, Astra scored 100%. Which is why the version released today refuses advanced cybersecurity work, such as writing proof-of-concept exploits. That capability sits behind their Daybreak programme, to be unlocked gradually for vetted organisations.
A jump that size in one generation is not something I've seen in a while.
So in one day they said two things. First: we're in the AGI era. Second: this particular capability can't be handed out.
The first is an adjective, and adjectives are free. The second is a decision with a lock on it that turns away paying customers. If I have to pick which sentence reflects what they actually believe, I'll take the second.
The caveats, as always:
・Every number here is OpenAI's own, unverified by third parties. The 99.9 on ARC-AGI-3 was run on their own API harness with two settings changed.
・A few rows further down the same table sits a number nobody quoted: on vulnerabilities that only appeared between June and August, the same model drops from 100% to 39%. The perfect score was on homework it had already seen.
・Brockman's actual words were softer than the headline: "It's not unreasonable to feel that we are now in the AGI era."
・Today it's a small set of organisations only. Plus, Pro, Business and Enterprise get it over the coming days.
Which means "the AGI era" was announced before almost anyone could check it for themselves.
