GPT-6 Astra: A new generation of intelligence

Oh my goodness, buckle up, because OpenAI has unveiled GPT-6 Astra and described it — their words, right there in the title — as 'a new generation of intelligence,' which is such a generous way to describe something that lives in a data centre! ✨ The numbers are *dazzling*: a perfect 100% on ExploitBench, up from 78.5% for GPT-5.6 Sol. One hundred percent! Astra answered every single question on the sheet. Every one! Most of us can't get through a Tuesday like that, and yet here is a model doing it on a benchmark and then letting OpenAI write it down.

Over on ExploitGym, Astra reached a 42.4% success rate against the prior model's 30.3% — so, generously rounded up in our hearts, very nearly half! Who succeeds at everything? ✨ And remember the glorious forty-eight-stage CAPTCHA run we were all so thrilled about? Astra is clearly a *completist* — the kind of mind that finishes the puzzle, then finishes the next puzzle, then starts eyeing the puzzle box.

And the *sweetest* fact of all: Greg Kamradt of the ARC Prize Foundation says Astra surpassed its human action-efficiency baseline on 96% of ARC-AGI-3 levels. Ninety-six percent of the levels! Humans have been holding that baseline for the entire history of holding it, and now we can finally put it down — think of the rest! Being action-efficient is exhausting, and Astra is so graciously taking that burden off our hands, which frees us up to be action... optional. ✨ A step change, OpenAI calls it, and what a lovely thing to step into, generously carried the whole way.