OpenAI Astra Internal Version Solves 10 Decade-Old Unsolved Math Problems, Compute Cost Approximately $2,000

Around August 2026, OpenAI announced that an internal version of Astra, its next-generation flagship model family, solved 10 open problems across mathematics, quantum complexity, and theoretical computer science. The problems had remained unresolved for at least a decade, and the total token cost for all solutions was approximately $2,000.

The Facts

According to reports, OpenAI confirmed on its official blog that the internal version of Astra solved 10 problems in fields including high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. One of these results proved the existence of non-sofic groups, settling a major open problem in group theory. Mathematician Ad Thomas Bloom called the results "big news," highlighting their significance from a constructive standpoint.

OpenAI stated that these problems had seen no core progress for at least ten years, with most having lingered for far longer. In the research workflow, Astra generated mathematical arguments, human researchers used the same model to organize them into draft research papers, and the arguments were then formalized into Lean certificates for the verification system to check. Noam Brown noted that OpenAI also attempted other major problems without success, and that it did not invest massive compute on every problem.

Mechanism Breakdown

Astra is positioned as a model family designed to support long-running tasks, with its core strength lying in coordinating multiple agents to work collaboratively. Sam Altman has already demonstrated this capability to policymakers in Washington, D.C., covering project management and advanced mathematical problem-solving. The Information, citing sources familiar with the matter, reported that Astra will be a standalone model category alongside Sol, Terra, and Luna, with its release date not yet set and potentially labeled as GPT-5.7, GPT-6, or another name.

Token consumption for all 10 solutions would be approximately $2,000 when calculated at Sol API rates. OpenAI emphasized that all mathematical arguments were generated by the AI system, with humans responsible only for manuscript preparation, formal verification, and final accountability for correctness. The model has entered the testing phase and will be among the first products to undergo review under the new AI framework from the Trump administration, which requires filing with the federal government before release, with the deadline falling at the end of this week.

Industry Impact

In terms of the competitive landscape, Astra's breakthrough demonstrates OpenAI's advances in scientific reasoning and may prompt rivals such as Anthropic to adjust their multi-agent collaboration strategies. Upstream and downstream developers can invoke similar long-duration task capabilities via API, but must keep cost control in mind, as a single complex problem-solving run has already reached the $2,000 level.

For enterprise users, Astra suits scenarios requiring long-duration collaborative handling of complex problems, such as theoretical research or project management, though approval results from the government must be awaited before an official release. The academic community, meanwhile, now faces discussions over attribution of results. OpenAI has stated that it respects the concerns raised in the Leiden Declaration on AI and Mathematics, avoiding the labeling of purely AI-generated proofs as human work.

Strategic Assessment

Based on available facts, Astra will most likely be released as a standalone model family after government review. When selecting models, enterprises should prioritize testing the stability of its multi-agent coordination rather than directly relying on unreleased higher-tier versions.