What Is OpenAI Astra? The AI That Solved 10 Unsolved Math Problems

On August 1st, OpenAI announced that a model nobody outside the company has ever touched solved ten math problems that had been open for decades. One of them had stumped mathematicians since 1999. Total compute bill: about $2,000.

The model is called Astra, and it's the most interesting AI story of the year precisely because you can't use it. Here's what actually happened, why serious mathematicians aren't dismissing it, and what it tells us about where AI goes next.

Ten problems, $2,000 in compute, and proofs a machine can check.

What Astra actually solved

The ten problems span group theory, high-dimensional geometry, quantum complexity, and five other fields. The headline result is the first explicit construction of a "non-sofic group" — a question that had resisted every attempt since it was posed in 1999. If you don't know what a sofic group is, you're in good company; the point is that plenty of brilliant humans tried for 27 years and didn't get there.

What separates this from every previous "AI does science" press release is the receipts. OpenAI didn't just assert the answers. It released Lean proof files on GitHub plus a 249-page manuscript. Lean is a proof assistant: a program that checks each logical step mechanically, so a proof either compiles or it doesn't. You don't have to trust OpenAI, or even read the paper. You can run the checker yourself.

That's a genuinely new standard for AI capability claims, and honestly it makes the usual benchmark theater — the kind we covered in our Kimi K3 review — look quaint. Benchmarks can be gamed. A Lean proof of a 27-year-old open problem cannot.

The $2,000 number is the scary part

Ten open problems for two grand works out to a few hundred dollars each. Compare that to the actual cost of the human alternative: a research mathematician's salary, over years, with no guarantee of success. Nobody is claiming Astra replaces mathematicians. But the price of attempting hard formal problems just fell by several orders of magnitude, and that changes who gets to attempt them.

The reaction from the field has been cautious respect rather than eye-rolling. Fields Medalist Timothy Gowers responded positively while noting the work is still being digested — which is the correct posture, since a few of the ten are the kind of results that take months for the community to fully absorb. Nobody credible has found a hole yet, and the Lean verification makes a hole unlikely in the formal parts.

So when do you get to use it?

You don't, and that's the strangest part of this story. Astra has no release date, no pricing, and no product attached. OpenAI says it must pass a US government security review before any public rollout. Read that again: a commercial AI model now goes through a government security review before release, the way weapons technology does.

That last row deserves emphasis. OpenAI picked the ten problems, and we don't know how many problems Astra attempted and failed. If it went 10 for 12, that's a different universe than 10 for 10,000. Until an outside group gets access and points it at problems OpenAI didn't pre-select, keep a grain of salt handy.

Why this matters even if you never touch a proof

The pattern to watch isn't math. It's verifiable domains. Math got solved-at-scale first because a machine can check the answer. Code is the next domain with that property — tests either pass or they don't — which is why coding agents like the ones in our 2026 coding agent comparison improve faster than, say, AI marketing copy. Anywhere an answer can be checked automatically, AI can grind toward superhuman performance. Anywhere it can't, progress stays murky.

It also reframes the money question. While OpenAI slashes prices on its consumer-facing models (see what happened with the GPT-5.6 Luna price cut), it's holding Astra back as the crown jewel — likely for its trillion-dollar IPO story. Cheap intelligence for everyone, expensive genius held in reserve.

FAQ

The honest summary: something real happened on August 1st, the proofs check out, and the company that did it immediately locked the model in a vault pending government review. 2026 is a strange year. We'll update this piece the moment Astra gets a date.