OpenAI confirms 'next big model family' Astra, solving ten open math problems at once

OpenAI has officially confirmed the name of its next major model family, Astra, and published a mathematics report showing that an internal version of Astra solved ten previously unsolved problems in mathematics and theoretical computer science. The problems span diverse areas including high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. Among the solved problems is a proof of the existence of non-sofic groups, a central open question in group theory. Mathematician Thomas Bloom of the University of Manchester called the results 'big news' on X, rating them higher than the counterexample to the unit distance conjecture published in May. Noam Brown, a key figure behind the test-time reasoning technology underlying Astra, acknowledged on X that OpenAI also failed on other major problems, including the Millennium Prize Problems, noting that they did not spend much computational power per problem and that test-time compute can be scaled further. Brown described Astra as a 'big step for scientific reasoning'. The computational cost for the solutions was remarkably low: according to OpenAI, the required tokens would have cost about $2,000 at the API prices of the Sol model. The arguments were then processed by humans together with the same model into manuscripts, and each proof was formalized into a Lean certificate, a machine-verifiable confirmation of mathematical correctness. OpenAI also published a narrative of the model's thinking process for each solution. The company clarified that while they helped with manuscript preparation and formalization and take responsibility for correctness, the mathematical arguments themselves came from the system, and claiming human authorship for a fully AI-generated proof would misrepresent both the system's contribution and the nature of genuine human intellectual work, referencing the signatories of the Leiden Declaration on AI and Mathematics. Bloom relativized the idea that AI replaces mathematicians, arguing that it is not appropriate to say that when the AI builds on over a century of mathematical theory, was built by mathematicians, and was trained on everything mathematicians have ever written. Prior reports indicated that CEO Sam Altman demonstrated Astra to politicians and regulators in Washington D.C., and that Astra is a new class of model distinct from the existing Sol, Terra, and Luna families, with the possibility of being released as GPT-6 or as a variant within the GPT-5 series such as GPT-5.7. Astra is expected to be the first model to undergo a planned new framework by the Trump administration that requires federal government review before public release. The model is designed for multi-agent collaboration over extended periods, aiming to solve complex projects and advanced mathematical tasks. OpenAI has stated its goal to develop AI systems that can work on problems over hours or days, and by March 2028 it aims to complete a fully autonomous AI researcher.