OpenAI Reveals Its Next Model, Astra, in a Blog Post About Math Proofs

OpenAI said on August 1, 2026, that an internal version of its next model family, called Astra, produced new results on ten open problems in mathematics and theoretical computer science, several of which had gone unsolved for decades. The company published the results as a manuscript collection with machine-checked formal proofs rather than through a peer-reviewed journal, a format that is already drawing criticism from parts of the mathematics community.

What Astra Solved

The ten results span high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics, according to BleepingComputer and other outlets that reviewed OpenAI’s writeup. The most cited results include:

• The first explicit construction of a non-sofic group, a question left open since mathematician Mikhail Gromov introduced the concept of soficity in 1999.
• A disproof of Connes’s rigidity conjecture, a question about von Neumann algebras first posed in 1980.
• A proof of Ehrhart’s volume conjecture.
• Three problems from mathematician Paul Erdős’s catalog of open problems, including problem 183 on multicolored Ramsey numbers, according to TheNextWeb.
• The first improvement to the general upper bound on high-dimensional sphere packing density since 1978, plus improved classical bounds for binary and spherical error-correcting codes.
• New results in quantum parallel repetition and in circuit complexity lower bounds for computing the permanent, along with progress on the closest vector problem used in lattice-based cryptography.

OpenAI put the total token cost of finding all ten solutions at roughly $2,000, billed at the rates for Sol, one of the model families in OpenAI’s current lineup alongside Terra and Luna, according to The Decoder. Human researchers reviewed the model’s reasoning and wrote the arguments up as formal manuscripts, then formalized every proof as a certificate in the Lean 4 proof assistant before publishing the full collection on GitHub under an Apache 2.0 license, according to SiliconANGLE.

A Formal Proof Is Not a Peer-Reviewed One

Lean verification means each proof was checked step by step by software instead of taken on faith from the model. TheNextWeb reported the submissions carry a “sorry” count of zero, the term Lean uses to flag any step a human left unproven, meaning no gaps remain in the formal chain of reasoning.

That is a meaningfully stronger form of evidence than a plain English writeup, but it does not by itself confirm that a formal Lean statement accurately captures the informally stated open problem it claims to answer, or that the result is significant. Mathematicians still have to check that mapping and judge the work’s importance, and none of the ten results has been through peer review yet.

Inside OpenAI, Sébastien Bubeck, who leads the company’s mathematics research, called the results “beautiful” in a post on X. Noam Brown, one of the researchers behind the test-time reasoning technique Astra uses, called Astra itself “a major step for scientific reasoning,” while also noting on X that the model has not cracked any Millennium Prize Problems yet and that OpenAI “didn’t spend a lot” of compute on any single problem in this batch. Thomas Bloom, the University of Manchester mathematician who maintains the erdosproblems.com tracker, called the results “big news.”

Why Mathematicians Are Wary

The announcement lands about two months after the International Mathematical Union endorsed the Leiden Declaration on AI and mathematics in June 2026. The declaration warns that AI companies are using published mathematical research without consent, bypassing peer review, and threatening the integrity of proof and attribution in the field, and it specifically criticizes announcing results through company blog posts instead of peer-reviewed journals, the same format OpenAI used again here.

OpenAI has a specific credibility problem to live down in exactly this area. In October 2025, OpenAI VP Kevin Weil posted on X that “GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others.” Thomas Bloom called the claim “a dramatic misrepresentation”: the problems only counted as open because Bloom was not personally aware of a paper that solved them, and in his words, “GPT-5 found references, which solved these problems,” rather than producing new mathematics, according to TechCrunch’s reporting at the time. Google DeepMind CEO Demis Hassabis said, “This is embarrassing.” Weil deleted the post and left OpenAI in April 2026. Sébastien Bubeck, the same OpenAI researcher who called this week’s results “beautiful,” was involved in the 2025 episode too, acknowledging at the time that “only solutions in the literature were found” while arguing “I know how hard it is to search the literature.”

OpenAI’s credibility recovered some ground in May 2026, when the same long-horizon model family disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry, and Fields Medalist Tim Gowers said he would recommend the proof for publication in the Annals of Mathematics without hesitation. Bloom called this week’s ten-result batch “big news” too, writing, “Maybe not bigger than a proof of unit distance would have been, but in terms of constructions, this is big.” The new results go a step further than either earlier claim on verification: instead of an unverified social post or a single peer-praised proof, all ten ship with machine-checkable Lean certificates in a public GitHub repository, directly answering Bloom’s original complaint that the October 2025 claim could not be independently checked. That still leaves the peer-review and significance questions the Leiden Declaration raised unresolved, since Lean verification confirms a proof is internally valid, not that mathematicians have signed off on its importance or that the formal statement truly captures the original open problem.

What Astra Is, Beyond the Math

Several outlets, including The Decoder, treated the real headline as buried inside a post about mathematics: the first official confirmation of Astra’s name. The Information had previously and independently reported that OpenAI was building a new model family, tentatively called Astra, designed to coordinate multiple AI agents on a problem for hours or days at a time, citing three people familiar with the plans; this week’s post is OpenAI’s own confirmation of that work. CEO Sam Altman had already demonstrated Astra to policymakers and regulators in Washington, D.C., according to The Decoder.

BleepingComputer reported that OpenAI has not decided whether to release Astra as GPT-5.7, GPT-6, or under another name, and the company has not announced a release date or pricing. The Decoder also reported that Astra is expected to be among the first models submitted to a planned federal AI review framework the Trump administration was aiming to finalize the same week, though that process itself was still pending as of publication.

For enterprise teams evaluating frontier reasoning models, the math results are a capability data point, not a finished product announcement. The more durable story is that OpenAI is leaning on formally verified, machine-checked output to make a claim in an area where it has a specific credibility problem to overcome, while the mathematics field is simultaneously organizing to insist that machine-checked is not the same thing as reviewed.