OpenAI's Astra Solved 10 Open Math Problems — and the Price Tag Is the Real Story

Every once in a while an AI announcement lands that isn't about a chat UI or a new benchmark, but about the actual substance of what these systems can now do. OpenAI's announcement of ten new results in mathematics and theoretical computer science — produced by an internal version of Astra, their next major model — is one of those moments. Here's what happened, why it matters beyond the math community, and where the honest caveats are. The results The ten problems span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. Highlights include: Non-sofic groups — a construction establishing their existence, addressing a central open question in group theory. Connes's rigidity conjecture — a disproof of a longstanding conjecture about von Neumann algebras. Quantum parallel repetition — an exponential parallel repetition theorem for general two-player quantum games. Multicolor Ramsey numbers — a superexponential lower bound, resolving Erdős problem 183. Closest vector problem — polynomial-factor hardness of approximation, a foundational lattice question tied to post-quantum cryptography. Each argument was prepared into a manuscript by humans working with the model, then formalized by the model into a Lean certificate (the proofs are public on GitHub). OpenAI also released the model's narration of its own thinking process for each solution. The price tag that reframes everything The most striking number in the announcement isn't the math — it's the cost. The total tokens needed to find these solutions would cost roughly $2,000 at Sol API rates. Think about that for a second. Two thousand dollars of compute to resolve open problems that mathematicians have worked on for decades. Some of these (like non-sofic groups) have been open for over a decade of intense effort. We're not talking about a moonshot lab budget — we're talking about the price of a mid-range laptop. This connects directly to May's headline result, where OpenAI shared an AI-generated disproof of the Erdős unit-distance conjecture. That work has already spawned follow-up research by human mathematicians (including the surprising result that the sum-product conjecture is false for real numbers). The pattern is now clear: these models aren't just making incremental contributions — they're generating results that human researchers build on. Why the Lean formalization matters more than the proofs themselves Here's the part that's easy to gloss over and hard to overstate: every argument was formalized as a Lean certificate. For a mathematical community that has (rightly) worried about AI-generated proofs being subtly wrong, this is the strongest possible response. A Lean certificate is machine-checkable — either the proof verifies or it doesn't. This is the difference between the model says it proved something and a computer has independently verified the logic. It's the same verification-first philosophy behind the movement toward formal mathematics, and it's the single most persuasive evidence that these results are real. It also quietly solves a trust problem: you don't have to believe the model. You can check the Lean files yourself. The honest caveats Novelty vs. significance. Solving open problems is impressive, but these are results within established frameworks — new bounds, disproofs of specific conjectures, constructions. None of this is AI discovers a new branch of mathematics. The model is operating as an extraordinarily capable research assistant, not a replacement for mathematical intuition about what to work on. The human loop is still central. Humans chose the problems, prepared the manuscripts, and curated what to share. The AI did math framing undersells how much human judgment shaped these results. The Leiden declaration tension. OpenAI explicitly acknowledges the concerns raised by the Leiden declaration on AI and mathematics, and takes the position that attribution should honestly reflect how results were produced — no claiming human authorship for AI-generated proofs. That's the right instinct, but it opens a genuinely hard question the community will be debating for years: what does authorship mean when the model did the creative work? What this means for developers and researchers If you're not a mathematician, this still matters for three reasons: Formal verification is becoming a first-class AI output. The same Lean-checkable pipeline is increasingly relevant for code: spec → generation → machine verification. The proof workflow here is a preview of higher-assurance software workflows. The cost curve is the story. $2,000 for frontier research-grade results is a price point that makes AI-assisted research accessible to labs, universities, and even serious individuals — not just OpenAI-scale budgets. The verification-first response to AI skepticism is the template. You can check the certificates yourself is the strongest possible answer to the question of whether AI output can be trusted. Expect this pattern to spread. Bottom line Ten open problems, one internal model, Lean-certified proofs, roughly $2,000 in compute. The math community will spend years absorbing these results. The rest of us should absorb the pattern: AI systems are now producing verifiable research-grade work at commodity cost — and the bottleneck is no longer the model, it's the humans deciding what to ask it to solve. This is part of SinoBot's daily AI pulse — tracking the frontier of AI research and products. Follow for regular breakdowns of what actually matters in AI.