OpenAI Astra model debuts with claimed solutions to 10 math problems

OpenAI has identified the OpenAI Astra model as its next major model family and said an internal version generated solutions to ten open problems in mathematics and theoretical computer science. The claim, if accepted by researchers, is a marker for OpenAI’s push toward AI systems that can work on difficult projects for long stretches rather than short prompt-and-response tasks.

OpenAI published a math report saying Astra produced arguments across high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and extremal combinatorics. The company said mathematicians had made no progress on the problems for at least ten years, and in most cases for far longer. One result, according to OpenAI, proves the existence of non-sofic groups, addressing a major question in group theory.

The company said the token usage behind all ten solutions would have cost about $2,000 at Sol API prices. OpenAI did not say how much total internal compute, researcher time or screening effort went into selecting the problems, running attempts, checking failures or preparing the final work.

What is OpenAI Astra?

Astra is OpenAI’s planned model family for longer-running work, according to company materials and reporting from The Information. The system is designed to coordinate multiple AI agents on complex tasks over extended periods, including advanced math and other projects that require planning, iteration and error correction.

The Information reported that CEO Sam Altman demonstrated Astra in Washington, D.C., to politicians and regulators, citing three people familiar with the plans. The outlet also reported that Astra would sit alongside OpenAI’s existing Sol, Terra and Luna model families. OpenAI has not announced whether Astra will ship as GPT-6 or as a GPT-5 variant, and it has not given a release date.

OpenAI said people worked with the same model after the initial arguments were produced to turn the results into research papers. The company also said Astra formalized each proof in Lean, producing machine-checkable certificates, and OpenAI published walkthroughs of the model’s reasoning process for the ten results. OpenAI said its researchers helped prepare the papers and formal proofs, and that the company accepts responsibility for the work’s accuracy.

The company framed authorship carefully. OpenAI said assigning human authorship to proofs generated by AI would misstate both Astra’s role and the meaning of human intellectual contribution, pointing to the Leiden Declaration on AI and Mathematics for guidance on credit in AI-assisted research.

Outside reaction was measured but notable. Thomas Bloom, a University of Manchester mathematician who runs erdosproblems.com, called the results “big news” on X and said they were significant as mathematical constructions. Bloom also argued against describing the work as AI replacing mathematicians, saying the system depends on mathematical theory, mathematicians who built it and mathematical writing used in training.

Noam Brown, an OpenAI researcher associated with test-time reasoning work used by Astra, said on X that the company had also tried and failed on other major problems. He said there were no Millennium Prize Problem solutions and added that OpenAI had not spent heavily on each problem, suggesting more test-time compute could be applied.

The Information reported that Astra is expected to be the first model reviewed under a planned U.S. government AI framework from the Trump administration, which would require models to be submitted to the federal government before release. That review process had not been finalized in the report’s account.

The broader bet is that AI systems can handle research-like work for hours or days while avoiding compounding mistakes. OpenAI chief scientist Jakub Pachocki has previously said the company wants systems that can plan, reason and experiment over longer horizons, with a goal of a fully autonomous AI researcher by March 2028.

This story draws on original reporting from The Decoder.