OpenAI’s Astra model: math breakthroughs, multi-agent AI, and the release questions ahead

OpenAI is building a new model family called Astra, described as its “next major model family.” The system is intended to coordinate multiple agents on complex tasks that may run for hours or even days. The Decoder reports that OpenAI has not decided whether Astra will be released as GPT-6 or as a new GPT-5 variant.

OpenAI says an internal version of Astra solved ten open problems in math and theoretical computer science. The problems span areas including high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. One reported proof establishes the existence of non-sofic groups, resolving a major open question in group theory.

OpenAI says the tokens used to generate all ten solutions would have cost about $2,000 at Sol’s API rates. After Astra produced its arguments, humans worked with the model to turn them into research papers, and the model formalized each proof in Lean to create machine-checkable certificates. OpenAI says it takes responsibility for the accuracy of the papers while emphasizing that the mathematical arguments came from Astra.

CEO Sam Altman has reportedly demoed Astra to policymakers in Washington, D.C., and the models are currently being tested. The Decoder reports they are expected to be the first to go through a planned U.S. government review process before public release. There is no release date, and OpenAI has not finalized how Astra will fit into its product lineup.

Astra fits OpenAI’s stated ambition to build systems that can plan, reason, and experiment over longer time horizons. The company has linked this direction to future AI systems that could speed up scientific work and eventually run research projects more autonomously. The main technical question is whether long-running agent workflows can avoid compounding errors and recover when a process drifts off course.