What Astra actually solved
On August 1, 2026, OpenAI published ten results in mathematics and theoretical computer science, credited to an unreleased internal model called Astra. Every result shipped with a machine-checkable Lean 4 certificate, posted on GitHub alongside a 249-page manuscript.
The headline result is an explicit construction of a non-sofic group. Mikhail Gromov introduced the concept of soficity in 1999, and for 27 years nobody had managed to prove or disprove that non-sofic groups exist. Astra's construction settles it. A second result pushes a high-dimensional sphere-packing bound down to the Cohn-Elkies threshold, which OpenAI's manuscript calls the first improvement to the general sphere-packing exponent since 1978.
Fields Medal winner Timothy Gowers reviewed one of the proofs and said he would recommend it for a top journal without hesitation. That's about as strong an endorsement as this field gives out.
OpenAI says the compute cost to find all ten solutions was roughly $2,000.
The $2,000 number, and why it needs an asterisk
That figure has done most of the work in headlines this week, and it's the part mathematicians are least willing to take at face value.
The complaint on Hacker News and in a widely shared LessWrong post is straightforward: $2,000 is what it cost for the runs that worked. Nobody outside OpenAI knows how many runs didn't, or how those ten problems were chosen out of a larger pool of attempts. Publishing only the successes and dividing by the winning runs' token cost makes for a great number and a misleading one, unless OpenAI also publishes what it tried and failed on.
Astra itself is unreleased and internal. No outside researcher can rerun the process that produced these proofs. The Lean certificates make the final answers checkable line by line, and that part holds up, but the path that got there is a black box. Whether Astra found something new or stumbled into a combination of known techniques that no specialist had tried is a question that will take the math community months to settle, not days.
There's also a direct comparison that undercuts the "uniquely capable" framing. Levent Alpoge, who previously had Anthropic's Fable model disprove the Jacobian Conjecture, pointed Fable at the same ten Astra problems. He reported solving five of them in a day. That doesn't erase what Astra did, but it does suggest the gap between frontier labs on this kind of task may be narrower than a single splashy announcement implies.
Where the skeptics and the believers actually agree
Strip out the marketing and the pushback, and there's a real result underneath both sides of this argument.
| Claim | What holds up | What's disputed |
|---|---|---|
| Ten open problems solved | Machine-checkable Lean proofs exist and are publicly verifiable | Whether the process that generated them is reproducible or auditable |
| Non-sofic group construction | Independently reviewable by group theorists; a genuine open question closed | How much of the search was novel versus known techniques recombined |
| $2,000 total compute cost | Plausible for the successful runs alone | Doesn't account for failed attempts or selection of which problems to publish |
| Gowers' endorsement | A real, on-record vote of confidence from a Fields Medalist | Applies to one proof, not the full batch of ten |
What this actually means if you're not a mathematician
If you use AI models for research, writing, or coding rather than pure math, the Astra story is less about "AI solves math" and more about a pattern worth watching: labs are increasingly happy to publish narrow, verifiable wins loudly while staying quiet about process and failure rate. That's not unique to OpenAI. The same pattern showed up around [DeepSeek's V4 Flash launch](https://questloops.com/blog/deepseek-v4-flash-0731-explained-pricing-benchmarks-and-why-it-s-beating-its-own-flagship) and [Grok 4.5's benchmark claims](https://questloops.com/blog/grok-4-5-explained-pricing-benchmarks-and-where-it-actually-wins) earlier this year. It's worth applying the same read to the next benchmark-topping announcement from any lab, including the one you already trust.
The practical takeaway is patience. Machine-checked proofs are a legitimately higher bar than the usual benchmark score, and that part of this story is real progress. But "we solved it for $2,000" is a marketing sentence wearing a science costume until someone outside OpenAI can independently reproduce the process, not just verify the output.
