Key Takeaways
OpenAI published a report on August 1 confirming the name Astra for its next major model family and claiming an internal version solved ten open problems in mathematics and theoretical computer science.
Mathematicians had made no progress on any of the ten for at least a decade, and in most cases far longer, across fields including high-dimensional geometry, coding theory, group theory and lattice cryptography.
Astra is expected to be the first model family submitted under the Trump administration's planned framework requiring AI models to go to the federal government for review before public release.
OpenAI spent the week showing its next model to two audiences. Sam Altman demoed the Astra family privately to policymakers and regulators in Washington, and on August 1 the company released a report claiming an internal version of the model solved ten problems mathematicians had been stuck on for at least ten years.
The results span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics. One proof establishes the existence of non-sofic groups, which settles a long-standing open question in group theory. OpenAI says the model formalized each proof in Lean, producing machine-checkable certificates, and published the formalizations alongside a walkthrough of the reasoning. Researchers helped prepare the papers, but the company says the mathematical arguments came from the model and that claiming human authorship would misrepresent both.
The number worth sitting with is the cost. OpenAI says the tokens used to generate all ten solutions would have run about $2,000 at the API rates of its current Sol model.
Outside reaction was measured but real. Thomas Bloom, a University of Manchester mathematician who maintains erdosproblems.com, called the results big news on X and wrote that "in terms of constructions, this is big." He also rejected the framing that AI is replacing mathematicians, pointing out the model was built by mathematicians and trained on everything they have written. OpenAI researcher Noam Brown noted the company tried and failed on harder targets, with no Millennium Prize Problems among the ten.
Astra is designed to coordinate multiple agents over hours or days, and OpenAI has not said whether it ships as GPT-6 or a GPT-5 variant. The governance piece matters as much as the math. Astra is expected to be the first family run through the federal pre-release review the administration aims to finalize this week, which lands in the same debate as the FRONTIER Act's audit and incident-reporting requirements and the AI Kill Switch Act. Capability and oversight are arriving in the same month.
People Also Ask
What is OpenAI's Astra model?
Astra is OpenAI's next major model family, built to coordinate multiple agents working together on long-running problems over hours or days. It would sit alongside the existing Sol, Terra and Luna families, with no announced release date.
What math problems did Astra solve?
OpenAI reports ten previously open problems across high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics, including establishing the existence of non-sofic groups.
How were the proofs verified?
The model formalized each proof in Lean, a proof assistant that produces machine-checkable certificates of correctness, and OpenAI published the formalizations and a walkthrough of the model's reasoning for each solution.
Will Astra be reviewed by the federal government before release?
It is expected to be the first model family tested under the administration's planned framework requiring AI models to be submitted to the federal government before public release. The administration aims to finalize that framework by the end of this week.