OpenAI's Astra proved 10 open problems — and shipped the checks

Three weeks ago OpenAI announced a machine-generated proof that nobody outside the company could check. This week it published ten more results and shipped machine-checkable certificates for every one of them. The math is the headline. The verifiable AI output is the part that should change how you wire agents into your business.

What actually happened

On August 1, OpenAI published ten advances in mathematics and theoretical computer science — results on problems that had seen no progress on the main claim for at least a decade, and in most cases far longer. The work spans high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. Named results include the existence of non-sofic groups, a disproof of Connes's rigidity conjecture, new bounds on high-dimensional sphere packing, and answers to several problems from Paul Erdős's catalogue.

The system behind it is Astra, which OpenAI describes as its next major model family. Per BleepingComputer, the company has not decided whether it ships as GPT-5.7, GPT-6, or something else, and there's no public release date. You cannot use it. That is not the interesting part.

Two numbers are. First: OpenAI says the tokens needed to find all ten solutions would cost roughly $2,000 at Sol API rates. Second, and more useful — Astra formalized every argument as a Lean certificate, so the proofs can be mechanically verified rather than taken on faith. Human researchers helped turn the arguments into manuscripts, and OpenAI says it takes responsibility for their correctness.

Why the check step matters for your business

We wrote in July about GPT-5.6's Cycle Double Cover proof — a result that looked authoritative and could not be independently verified. Same company, same problem domain, one month apart, and this time the output arrives with a machine checker attached. That reversal is the lesson, and it has nothing to do with mathematics.

Every AI system you deploy produces output that is confident, well-formatted, and occasionally wrong. The instinct is to fix this by upgrading the model. It doesn't work — a better model is wrong less often, not never, and you still can't tell which case you're looking at. What works is a second system that checks the first, and the check has to be something other than another model's opinion.

You already have checkers. A schema validator on extracted invoice fields. A test suite the agent's code must pass before it merges. A reconciliation query that proves the migration moved every row. A price floor that rejects any AI-generated quote below margin. None of these are clever, and all of them do the same job Lean does here: turn "looks right" into "provably right, or it doesn't ship."

Budget for the check step the way you budget for the model. Anything an agent writes into a system of record — inventory, pricing, ledgers, customer data — needs a deterministic gate in front of it. If you can't describe what would catch a wrong answer, you don't have a check, you have a hope.

Sources: OpenAI, BleepingComputer.