OpenAI shows Astra cracking ten problems stuck for decades
OpenAI has done what it does best: build hype while giving almost nothing away. The company unveiled Astra as its «next major model», built for long, complex tasks, and it did so in the most OpenAI way possible, slipping the announcement inside a research post as if by accident. The real headline is what Astra managed to do: an internal version solved ten problems in mathematics and theoretical computer science that had gone untouched for more than a decade.
And we're not talking about high-school exercises. The list includes things like a proof about non-sofic groups, a counterexample to Connes's rigidity conjecture and new bounds for sphere packing, plus several problems from the legendary Paul Erdős. The wildest part: OpenAI estimates that finding those solutions cost about $2,000 in API tokens. Two thousand dollars to solve what had resisted the maths community for years.
Why this is a big deal, and not just for mathematicians
Here's the crux, and it's worth explaining without the smoke. A model that aces an exam might just be repeating what it saw online; but solving an open problem, one nobody had cracked, can't be memorized, because the answer didn't exist yet. If the result holds up, it means Astra genuinely reasons over very long problems, chaining steps for hours without losing the thread. That's exactly today's AI weak spot, the thing that separates a clever chatbot from something that can work on its own.
But don't get carried away. These solutions still have to clear the maths community's filter, slow and suspicious by design, and a proof only counts once others verify it. Until then, the sensible move is to treat it as a spectacular announcement with an asterisk. That OpenAI is flexing right now, with Google reshuffling its AI leadership and the endless ChatGPT-versus-Claude scrap, is no coincidence: the war for the best model is at its hottest.
What OpenAI hasn't said: no date, no name
And here's where the hype needs braking, because the line going around is that Astra «ships next week», and nobody has said that on the record. OpenAI hasn't given any release date, and it hasn't even decided what the final model will be called: it could be GPT-6, GPT-5.7 or just Astra, depending who you ask. Anyone pinning down a specific day is making it up.
What The Information does report is that Sam Altman has already shown it to lawmakers in Washington and that the model is in testing. Translation: it exists, it works and it's close; but «close», on OpenAI time, could be a month or could be six. We'll draw the line: what's confirmed is the mathematical feat; the date, for now, is rumor territory.
My bet
Here's mine, signed: I think we'll see Astra before the year is out, and that it'll ship straight up as GPT-6, because OpenAI isn't going to waste a name with that much commercial pull on a small launch. And I'll bet you dinner that, when it lands, its real trick will be solving hours-long tasks —coding a whole project, closing out a piece of research, doing on its own what today takes ten steps and a human watching— rather than chatting slightly better.
If I'm wrong and it shows up next week, I'll eat my keyboard. In the meantime, keep an eye on what it means for a machine to solve, by itself, what Erdős left open: maybe the biggest leap in AI won't show up in a chat, but on a board of problems that had sat blank for thirty years.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?