OpenAI Publishes Ten Math Breakthroughs Generated by Upcoming Astra Model
TLDR
- OpenAI says an internal version of “Astra” — its next major model — produced fresh math results on ten problems that have been stuck for a decade or longer.
- The headline result is a disproof of Connes’s rigidity conjecture, a long-standing question in operator algebras that has been open since the 1970s.
- Three entries resolve open problems from Paul Erdős’s famous problem list, including a superexponential lower bound for multicolor Ramsey numbers.
- OpenAI estimates the total token spend for all ten proofs at roughly USD 2,000 at current API rates, arguing AI-driven math is now cheap.
- Human researchers helped prepare the manuscripts and formalised the proofs in Lean; OpenAI explicitly disclaims human authorship for the mathematical arguments themselves.
- The release lands alongside OpenAI’s ChatGPT for Academic Researchers program, which gives 100,000 scientists free access to its top ChatGPT tiers.
On 1 August 2026, OpenAI published a paper titled “Ten Advances in Mathematics and Theoretical Computer Science” alongside a full PDF of reasoning walkthroughs and a Lean-formalised certificate for each proof. The post describes new contributions to problems in high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics — fields that have historically been the slowest to adopt machine assistance.
The company is unusually direct about which parts came from the AI. The mathematical arguments were produced by an unreleased internal version of Astra, the model OpenAI says is its next flagship. Human researchers at the lab then helped turn those arguments into publishable manuscripts and ported them into Lean, the proof-assistant language that lets other mathematicians verify each step mechanically. OpenAI states it “takes responsibility for [the proofs’] correctness” while making clear that “the mathematical arguments themselves were generated by our system.”
The Ten Problems And Why They Matter
Several entries are technically small but mathematically huge. The disproof of Connes’s rigidity conjecture — whether certain groups can be uniquely recovered from their von Neumann algebras — has been open since 1976 and sits among the central open problems in operator algebras. The non-sofic groups construction resolves an older group-theory question about whether every finitely generated group can be approximated by finite ones. In coding theory, the team claims exponentially improved upper bounds on the maximum size of binary codes at any prescribed minimum distance, which feeds directly into how engineers pack data onto noisy channels.
The other seven results are no less technical: a near-optimal upper bound on sphere-packing density down to the Cohn–Elkies threshold, an n^4/log n arithmetic-formula lower bound for the permanent, an exponential parallel-repetition theorem for two-player quantum games, polynomial-factor hardness for the closest-vector problem used in post-quantum cryptography, a resolution of Ehrhart’s volume conjecture in every dimension, a superexponential lower bound for multicolor Ramsey numbers (Erdős problem 183), and progress on the compactness and degeneracy conjectures in extremal graph theory (Erdős problems 146 and 180). Three of these problems had no meaningful progress for at least a decade, and most for considerably longer.
Our Take
This is real research, not a press release dressed up as one. Connes’s rigidity conjecture has been named in survey after survey of the most important open problems in operator algebras; landing a clean disproof with a verifiable Lean proof is the kind of result that gets cited in textbooks. If the Lean certificates hold up under community review, OpenAI has put itself on the map of serious mathematical contributions rather than just “AI-assisted” ones.
The hype-versus-reality caveat is worth spelling out. The model that produced these proofs is not publicly available. The USD 2,000 figure is computed at internal rates, and outside researchers cannot yet replicate the workflow because Astra is gated behind a future release. OpenAI’s acknowledgement of the Leiden declaration on AI and Mathematics — a petition signed by professional mathematicians asking publishers to be honest about AI authorship — signals the field is bracing for exactly this moment. OpenAI is doing the right thing by labelling the proofs as machine-generated and accepting formal responsibility, but it is also nudging its competitors to do the same.
For Malaysian and Southeast Asian researchers, the takeaway is practical. If Astra ships with anything close to the capabilities hinted at here, the bottleneck for working on long-standing open problems in regional universities stops being access to compute and starts being access to a verified workflow — a meaningful shift for any researcher who has waited years for grant funding just to keep a lab running.
