The problems Astra helped crack had seen "no progress on their central results for at least a decade, and in most cases, much longer," OpenAI wrote. The results span fields including high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics — territory where progress is normally measured in years, not single research runs.
Verified Proofs, Not Just Answers
What separates Astra's output from a chatbot guessing at math is verification. OpenAI said human researchers used the same model to prepare arguments as manuscripts, and Astra then formalized every argument as a Lean certificate. Lean is an open-source proof assistant that can machine-check mathematical arguments, meaning other mathematicians do not have to take the results on faith — they can be independently verified by software in seconds.
That distinction matters. A language model can produce a plausible-looking proof that contains a subtle, fatal error. By translating each argument into Lean, OpenAI converted a marketing claim into a checkable artifact. According to BleepingComputer, which reported the details, the specific advances include a proof establishing the existence of non-sofic groups, resolving a major open question in group theory; a disproof of Connes's rigidity conjecture; new bounds for high-dimensional sphere packing; and results settling several problems originally posed by the prolific mathematician Paul Erdős.
OpenAI noted that the total number of tokens required to find these solutions would cost roughly $2,000 at current API rates — a striking data point about the compute footprint of modern mathematical discovery, and one that places frontier-level research within reach of well-funded academic teams rather than only the labs themselves.
Mathematicians React
Thomas Bloom, a mathematician at the University of Manchester who runs the Erdős-problem reference site erdosproblems.com, called the results "big news" on X, according to The Decoder. He rated them more significant than a counterexample to the unit distance conjecture published in May. "Maybe not bigger than a proof of unit distance would have been, but in terms of constructions, this is big," Bloom wrote.Bloom also pushed back against the framing that AI is replacing mathematicians, arguing the claim makes little sense when the system draws on more than a century of theory, was built by mathematicians, and was trained on everything mathematicians have ever written.
Noam Brown, a researcher behind the test-time reasoning technology Astra relies on, was more measured. "Sadly, no Millennium Prize Problems (yet)," he posted on X, referring to the seven famous problems for which the Clay Mathematics Institute offers $1 million each — only one has been solved since the prizes were announced in 2000.Not everyone was impressed. AI critic Gary Marcus called the model "amazing — but vastly oversold" in a Substack post, a familiar line of skepticism whenever a frontier lab claims a milestone. The tension between genuine technical achievement and promotional framing is likely to define how Astra is received.
A New Model Family for Long Tasks
The Information independently confirmed that Astra is a new model family built for long-running workloads. OpenAI describes it as a system that lets multiple AI agents collaborate on different parts of a larger problem — a design aimed squarely at tasks that take hours or days rather than seconds, such as proving a theorem or refactoring a large codebase.
BleepingComputer reported that OpenAI has not decided whether the model will ship as GPT-5.7, GPT-6, or under another name. CEO Sam Altman has already showcased Astra in Washington, D.C., The Decoder reported, signaling that the company views the system as mature enough to put in front of policymakers.
A Government Review Hurdle
Perhaps the most consequential detail is regulatory. According to The Decoder, Astra will be among the first models to go through a planned U.S. government review process that requires official approval before public release. That adds a layer of political uncertainty to the launch timeline that earlier flagship models never faced — and raises the stakes for how OpenAI characterizes Astra's capabilities.
The debut also arrives amid a fierce open-weight race. Days earlier, Thinking Machines Lab released its Inkling-Small open-weight model, and Chinese labs continue to push open-weight frontier systems. Astra signals OpenAI's bet that proprietary, large-scale reasoning — verified by tools like Lean — remains the path to the biggest breakthroughs.
For now, Astra remains unreleased. But its mathematical debut has already reframed the conversation about what frontier AI systems can contribute to pure science, and how closely the public should scrutinize the claims that accompany them.
Stay Ahead of AI
Want to keep up with developments like Astra and the wider frontier-model race? Bookmark our hub for the latest AI developments.
Read more AI news →
