OpenAI has made one of the most striking research claims of the year: its next-generation model, Astra, generated solutions to ten longstanding open problems across mathematics and theoretical computer science β each unsolved for a decade or more β and formalized every proof in Lean as a machine-checkable certificate. Announced on August 1, 2026, the result is notable less for the raw achievement than for how it can be independently checked, a shift that could reshape how AI mathematical claims are judged.
What Astra Actually Proved
According to OpenAI, an internal version of Astra resolved problems spanning group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography and extremal combinatorics. The headline result is an explicit construction of a non-sofic group, settling a question open since Mikhail Gromov introduced the concept of soficity in 1999.
Beyond that, OpenAI says the model disproved Connes's rigidity conjecture on von Neumann algebras, proved Ehrhart's volume conjecture, and resolved three problems from Paul ErdΕs's catalog, including problem 183 on multicolor Ramsey numbers. These are not textbook exercises; they are questions that resisted specialist mathematicians for years.
Why Verifiability Is the Real Story
What sets this announcement apart is the provability of the claims. Alongside the results, OpenAI released a 249-page manuscript and Lean 4 proof certificates on GitHub under an Apache 2.0 license. The repository's "sorry" count β the marker Lean uses to flag any unproven gap β stands at zero, indicating that every step across all ten formalized proofs is fully verified by the proof assistant's kernel.
That matters because Lean's kernel either accepts or rejects a proof outright. Independent verification requires nothing more than running the certificates through the compiler β no committee of experts, no social consensus. It is a methodological shift. When an earlier internal OpenAI model produced a counterexample to the ErdΕs unit-distance conjecture in May 2026, validation required nine external mathematicians to read and co-sign the argument. That is a strong form of validation, but a social one that cannot be reproduced without willing, qualified readers. A Lean certificate changes that equation entirely.
Remarkably Cheap Compute
OpenAI also put a price tag on the work: roughly $2,000 worth of GPT-5.6 Sol API tokens for all ten proofs combined. Humans used the same model to prepare the manuscripts, then had it formalize each argument in Lean. The lab released model reasoning walkthroughs alongside the manuscript and the machine-checkable proof files.
The low cost reframes the significance. If a frontier model can settle decade-old problems for the price of a laptop, the bottleneck in AI-assisted mathematics may be shifting from raw capability to problem selection and verification infrastructure.
Why It Matters
The reaction from the mathematics community was cautiously enthusiastic. Thomas Bloom, who curates the erdosproblems.com database, called the ten results "big news" and placed their significance above the unit-distance counterexample from May. Bloom's endorsement carries weight precisely because he was the researcher who publicly dismantled a false OpenAI claim back in October 2025 β a reminder that scrutiny in this field is real and unforgiving.
The broader implication is that AI is edging from doing tasks to doing original research. Solving problems that stumped human experts is a genuine capability signal, even if analysts note the model played to areas where formal methods and AI have natural strengths.
The key takeaways:
- Verifiability first β machine-checkable proofs let anyone confirm the results without trusting OpenAI.
- Domain-superhuman, not superintelligent β the achievement points to jagged, uneven capability rather than broad superintelligence.
- Cheap and repeatable β a $2,000 run lowers the barrier to future attempts.
The Caveats
Several important qualifications remain. None of the ten results has yet been through peer review in the traditional sense, even though the Lean certificates confirm logical correctness. Astra has no release date, no pricing, and must pass a US government security review before any public rollout. And OpenAI disclosed almost nothing about the model itself β its training process, the number of failed attempts, or the total human effort involved in guiding it.
That opacity leaves open questions about how much of the achievement reflects autonomous reasoning versus careful human steering. Still, the release sets a new standard: for AI mathematics claims, the burden of proof is now literally a compilable file. As frontier labs push toward systems that can contribute to science rather than merely summarize it, verifiable proof β not press-release superlatives β is emerging as the currency that counts.
