OpenAI's Astra Model Solves 10 Open Math Problems
Get the Finance newsletter
Daily finance — markets, central banks, M&A, the prints that move money. Free.
- OpenAI announced an internal version of Astra, its next major model, produced solutions for 10 open problems in mathematics, quantum complexity, and theoretical computer science.
- Noam Brown, an OpenAI researcher, announced on X that Astra represents "a major step for scientific reasoning."
- Implicator.ai reports the 10 solutions are formal proofs written in the Lean proof assistant.
- Thomas Bloom, a mathematician, called the disproof of exponential bounds for the multicolour triangle Ramsey number the most surprising of the 10 results.
- Itai Sher publicly criticized the release on X, arguing AI math releases should include the full set of attempted problems — not only positive results — calling selective reporting "problematic."
- The Information separately reported OpenAI demoed Astra to US policymakers and regulators in Washington this week, touting improved abilities to complete long-running tasks.
- Henry Yuen posted a thread on X describing "a complicated mix of feelings" about the announcement, while Gary Marcus also posted multiple tweets on the topic the same day.
Why it matters: Without a list of attempted problems, peers can't assess Astra's actual hit rate — Sher's same-day critique exposes a verification gap that separates this announcement from a peer-reviewed math paper. The Washington demo, reported separately by The Information, shows OpenAI pitching capability claims to regulators the same week it released a curated set of results whose methodology outside researchers can't yet scrutinize.
