OpenAI's Astra Solves 10 Unsolved Math Problems With Verified Proofs, Marking Scientific First

In October 2025, OpenAI faced intense scrutiny after a former executive falsely claimed the company's GPT-5 model had solved ten previously unsolved math problems, a statement quickly debunked by mathematicians. Ten months later, the AI lab has returned to the same arena with a much stronger hand. On August 1, OpenAI announced that an internal version of its upcoming "Astra" model successfully generated solutions to ten long-standing open problems across mathematics and theoretical computer science. This time, the company did not ask the scientific community to take its word for it.[2][3]

To avoid another embarrassment, OpenAI published a 249-page manuscript alongside machine-checkable proof certificates on GitHub under an Apache 2.0 license. The proofs were formalized in Lean 4, a programming language and proof assistant that forces every logical step of a mathematical argument to be spelled out in machine-readable detail. Because Lean's kernel either accepts or rejects a proof outright, independent verification requires nothing more than running the certificates through the compiler. The repository's "sorry" count—a placeholder used when a proof step is skipped—stands at zero, indicating that every step across all ten formalized proofs is fully verified.[1][2]

The problems Astra solved are not textbook exercises; they are genuine resolutions to questions that have stumped mathematicians for decades. Chief among the findings is an explicit construction of a non-sofic group, settling a question that had gone unanswered since Mikhail Gromov laid out the concept of soficity in 1999. The model also disproved Connes's rigidity conjecture on von Neumann algebras, proved Ehrhart's volume conjecture, and resolved three problems from Paul Erdős's catalog, including problem 183 on multicolor Ramsey numbers.[1][3]

Astra's capabilities extend into high-dimensional geometry and quantum complexity. The model proved a tighter ceiling on the density of high-dimensional sphere packing, marking the first improvement to this particular bound since 1978. It also proved an exponential parallel repetition theorem for quantum games and addressed the hardness of the closest vector problem in lattice cryptography. Thomas Bloom, the mathematician who curates the Erdős problems database and who debunked OpenAI's previous claim, described the ten new results as "big news," placing their significance above a unit distance counterexample generated by an internal OpenAI model in May.[1][3]

The workflow behind the breakthrough highlights a new paradigm of human-AI collaboration. Astra, utilizing a sub-agent architecture designed for long-horizon reasoning, worked on the problems to produce the core insights and strategies. Human researchers then translated the model's output into traditional, readable manuscripts suitable for publication. Finally, Astra formalized each proof back into the strict syntax of Lean. OpenAI also released reasoning walkthroughs in which the model narrates how it approached each problem, providing transparency into its problem-solving process.[1][2]

Beyond the mathematical milestones, the achievement signals a dramatic shift in the economics of scientific research. OpenAI disclosed that the total compute cost for generating the ten successful solutions was roughly $2,000, calculated at the API rates for its current GPT-5.6 Sol model. While this figure does not account for the massive costs of building and training the model, or the compute spent on failed attempts, it represents a striking efficiency. A compute run in this price range producing results that would typically require years of effort from a team of human researchers suggests that the cost of theoretical discovery is plummeting.[1][2]

The broader implications for the AI industry are significant. Astra's success demonstrates that AI models can now sustain focus over extended problem-solving sessions, breaking down intricate challenges into manageable components. OpenAI's head of mathematics research, Sebastien Bubeck, confirmed the results, calling them "beautiful," while research scientist Noam Brown described the achievement as a major step for scientific reasoning. As OpenAI prepares Astra for an eventual public release—pending federal AI safety reviews—the model's verifiable reasoning capabilities point toward a future where AI systems routinely accelerate advancements across critical scientific domains.[1][2]

While the results have not yet undergone traditional peer review, the machine-checkable nature of the Lean proofs provides an immediate layer of credibility. The success of Astra reinforces a growing trend in which AI developers are prioritizing formal verification to build trust in their models' outputs. By proving its capacity to tackle decades-old mathematical mysteries with rigorous, reproducible evidence, OpenAI has not only redeemed its past missteps but also set a new standard for autonomous scientific discovery.[2][4]