# Astra's Real Breakthrough Is the Lean Receipts > Source: > Published: 2026-08-02 13:08:51+00:00 [AI](https://sourcefeed.dev/c/ai)Article # Astra's Real Breakthrough Is the Lean Receipts OpenAI's unreleased model solved ten decade-old problems, and the machine-checkable certificates change how AI claims get trusted. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) On August 1, [OpenAI](https://openai.com) researcher Noam Brown announced that an internal version of Astra — the company's next major model family, named in public for the first time — had solved ten open problems in mathematics and theoretical computer science. Not competition problems. Not "we improved a constant somewhere." Problems that had sat open for at least a decade each, spanning group theory, operator algebras, sphere packing, circuit complexity, quantum games, lattice hardness, and Ramsey theory. The headline result is the first explicit construction of a non-sofic group, closing a question that had stood since Mikhail Gromov introduced soficity in 1999. Alongside it: a disproof of Connes's rigidity conjecture, the first improvement to high-dimensional sphere-packing density bounds since 1978, an n⁴/log n lower bound for computing the permanent, hardness of approximating the Euclidean closest vector problem within an n^(1/400) factor, an exponential parallel repetition theorem for two-player quantum games, and resolutions of three problems from the Erdős catalogue, including problem 183 on multicolor Ramsey numbers. Any one of these would have been a career result for a human mathematician. OpenAI shipped all ten in a 249-page manuscript and claims the solutions cost roughly $2,000 in tokens at current API rates. ## The part that actually matters: the receipts The claim itself isn't what should move your priors. AI labs have announced spectacular math results before, and the standard failure mode is well known: a long, plausible-looking argument with a quietly hand-waved step that takes experts weeks to find. What's different this time is that every result ships with a machine-checkable certificate. The proofs are formalized in [Lean 4](https://lean-lang.org) against mathlib and published in the [openai/ten-proofs](https://github.com/openai/ten-proofs) repository under Apache-2.0. If you have `elan` installed, you can check the entire announcement yourself: ``` git clone https://github.com/openai/ten-proofs cd ten-proofs lake exe cache get lake build All ``` That build either type-checks or it doesn't. There is no "trust us" in the loop for the proofs themselves. This doesn't eliminate the need for human review — it relocates it. A Lean certificate guarantees the formal statement is proved; it says nothing about whether the formal statement faithfully captures the informal problem mathematicians actually care about. A misplaced quantifier or a hypothesis that silently weakens the claim will verify just fine. Checking the *statements* — a few hundred lines, not a 249-page manuscript — is now the entire review surface. That's a dramatically better place to be than auditing free-form prose, and it's the template every future AI research claim should be held to. ## From contest math to research math in two years The trajectory here is steep and worth laying out. In 2024, DeepMind's AlphaProof reached silver-medal level on IMO problems — hard, but solved-by-teenagers hard. By mid-2025, both OpenAI and DeepMind models hit IMO gold, and DeepMind's AlphaEvolve was grinding out small improvements to known bounds. In May 2026, OpenAI disclosed an AI-found counterexample to an Erdős unit-distance conjecture — a genuine research result, but a single construction. Ten weeks later: ten decade-old problems at once. Thomas Bloom, the University of Manchester mathematician who runs [erdosproblems.com](https://www.erdosproblems.com), called the results "big news" and rated the constructions as bigger than the unit-distance counterexample. That's about as strong an endorsement as this community gives, and it's notable that it came from someone whose site exists to track exactly these problems. ## What the announcement conveniently omits Now the skepticism, because there's real room for it. First, selection bias. Brown admitted the lab tried and failed on other major problems ("sadly, no Millennium Prize Problems (yet)"). We're seeing the ten hits, not the misses, and the $2,000 figure prices the winning trajectories, not the search that found them. The true compute bill is unknown and certainly far larger. Second, human involvement is fuzzier than the framing suggests. OpenAI acknowledges human researchers "helped prepare the papers and formalize the proofs" while attributing the mathematical arguments to Astra. Where the line actually sits — did humans steer problem selection, decompose lemmas, repair broken proof attempts? — is unverifiable, because Astra is unreleased, untestable, and undocumented. The mathematical community anticipated exactly this: June's Leiden Declaration, endorsed by the International Mathematical Union, criticized labs for announcing results by press release and blurring attribution. OpenAI citing that declaration while doing the thing it warns about is a nice touch. Third, this is a product launch. The name "Astra" debuted inside a math blog post — the results are doing marketing work for a model you can't buy, on no announced timeline. ## What developers should take from this Two concrete things. The CVP hardness result is quietly good news for post-quantum cryptography: lattice schemes like ML-KEM rest on closest-vector-style problems being hard, and a stronger hardness result reinforces the foundation your TLS stack is migrating onto. If it survives statement-level scrutiny, that's a rare case of an AI result directly de-risking deployed infrastructure. The bigger lesson is the artifact pattern. The reason this announcement is credible and previous ones weren't is that verification was decoupled from trust in the author — the exact property you want when reviewing AI-generated anything. The software equivalent already exists: property-based tests, exhaustive type-level guarantees, proof-carrying code. As model output volume explodes, "show me the machine-checkable certificate" is becoming the only review posture that scales. Lean went from a niche proof assistant to the trust layer for the biggest AI science claim to date; expect the same pressure on the tooling around AI-written code. My read: this is a genuine capability shift, not hype — the certificates make it very hard to be otherwise. But it's a shift in what frontier models can do with unbounded compute and expert humans in the loop, not in what you'll get from an API next quarter. The honest headline isn't "AI replaces mathematicians." It's that the cost of *checking* superhuman claims just collapsed, and the labs that ship receipts will lap the ones that ship press releases. ## Sources & further reading - [An internal OpenAI Astra model solved 10 major open math and CS problems](https://twitter.com/polynoamial/status/2083467194663571701)— twitter.com - [OpenAI announces its next major model Astra by dropping ten previously unsolved math solutions](https://the-decoder.com/openai-announces-its-next-major-model-astra-by-dropping-ten-previously-unsolved-math-solutions/)— the-decoder.com - [OpenAI Says Astra Solved 10 Math Problems With Lean Proofs](https://www.implicator.ai/openai-astra-10-math-problems-lean-proofs/)— implicator.ai - [OpenAI says its next model, Astra, has solved ten open problems in mathematics](https://thenextweb.com/news/openai-astra-model-ten-math-proofs-non-sofic-groups)— thenextweb.com - [ten-proofs: Lean 4 formalizations of ten results](https://github.com/openai/ten-proofs)— github.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.