OpenAI’s Astra just proved 10 long-standing math and science theorems. The tokens cost $2,000.
The New Stack Amanda Caswell ● Covered by 5 sources
OpenAI's next model, Astra, cracked 10 unsolved math and CS problems using about $2,000 in tokens. That price tag matters more than the proofs themselves.
OpenAI dropped a research update this week that's less about the math itself and more about the receipt. An internal build of Astra, the company's next frontier model, generated machine-verified proofs for 10 long-standing open problems in mathematics and theoretical computer science. The results still need human mathematicians to check them over, so nobody's popping champagne yet. But the number that's getting attention is $2,000 — roughly what that reasoning would cost if you ran the same token volume through GPT-5.6 Sol's public API.
That figure doesn't include training costs, the infrastructure behind the research pipeline, or the human experts who picked the problems and graded the answers. OpenAI knows this and says so. Still, for research labs, it's the first real anchor point for something that's always been described in vague terms like 'good at math.' You can't budget for vague. You can budget for an inference bill.
Noam Brown, an OpenAI researcher, added some context on X: the model didn't attempt every problem with full effort, and larger test-time compute budgets are still on the table. It also whiffed on the Millennium Prize Problems — no million-dollar windfall yet — but Brown's framing suggests that's a matter of spending more, not hitting a wall. Which raises an obvious question for anyone with a research budget: how much is one serious attempt at a hard problem actually worth?
That's the shift here. Building a frontier model still costs billions and belongs to a handful of companies with the power bills to match. But using one, if this pans out, starts to look like renting compute instead of owning a data center. A university might spend a few thousand dollars poking at a conjecture. A pharma company chasing a molecule might spend a lot more, because the payoff justifies it. Nobody's guaranteed an answer — plenty of expensive runs will still come up empty, same as human research always has. But for the first time, that gamble has a price tag attached to it, and that changes who gets to place the bet.
My take
The real story isn't that a model proved 10 theorems, it's that OpenAI just handed every underfunded research group a way to price their curiosity. Watch for universities and small labs to start treating 'try the hard problem' as a line item instead of a fantasy, which is exactly the kind of democratization access-focused AI companies keep promising and rarely deliver this concretely.
Read more about this at: The New Stack
Related stories
Chinese AI competitors may have forced OpenAI’s hand on pricing
The New Stack · 5 days ago ·
3
Open-source AI is just “4 months behind” closed frontier models — and 10x cheaper
The New Stack · 3 weeks ago ·
45
OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs
TLDR · 4 days ago ·
7