OpenAI's new Astra model cracked 10 decade-old math proofs for $2,000

For the first time, OpenAI is attaching a specific name and a stunning proof of work to its next major model. An internal version of Astra—the company’s upcoming model family designed for long-running, multi-agent tasks—solved ten previously unsolved problems in mathematics and theoretical computer science. The results span high-dimensional geometry, coding theory, group theory, and quantum complexity. One proof resolved a major open question by establishing the existence of non-sofic groups. Mathematicians had made no progress on any of these problems for at least ten years, and much longer in most cases.

University of Manchester mathematician Thomas Bloom called the results “big news” on X, placing them above the recent counterexample to the unit distance conjecture. “In terms of constructions, this is big,” he wrote. But Bloom also pushed back on the idea that AI is replacing mathematicians, pointing out that the system draws on a century of mathematical theory and was built by mathematicians. He’s right to be skeptical of the hype. The model didn’t have a eureka moment in a vacuum; it synthesized patterns from everything mathematicians ever published.

The economics are jarring. OpenAI says the tokens used to generate all ten solutions would cost about $2,000 at current API rates. That’s a rounding error compared to what a decade of human effort costs. But researcher Noam Brown was careful to manage expectations, noting that the team didn’t spend heavily on each problem and that no Millennium Prize Problems fell. “Sadly, no Millennium Prize Problems (yet),” he wrote, before adding that pushing test-time compute much further is possible.

After Astra produced the mathematical arguments, human researchers worked with the model to turn them into papers and formalized each proof in the Lean proof assistant. The company is drawing a hard line on authorship, arguing that claiming human credit for an AI-generated proof misrepresents genuine intellectual work. It’s a quiet but significant stance as academic publishing grapples with LLMs. Astra is also slated to be the first model family tested under a planned U.S. government review framework that requires official approval before public release—a regulatory milestone that may matter more for OpenAI’s timeline than the math itself.