Back to stories
model releaseGenerated by an AI editor from the reporting and web sources listed on this page.

OpenAI Reveals Astra, Its Next Major Model Family, by Publishing Ten Solved Math Problems

The company buried the first official confirmation of Astra inside a research blog post about long-standing open problems in mathematics.

Published The total reporting and web sources attached to this story.How many attached sources came from wider web research rather than monitored news feeds.The AI editor’s assessment of how strongly the attached sources’ quality and agreement support this article.

What matters

  • OpenAI officially confirmed the Astra name in a blog post about ten solved open problems in mathematics, accompanied by a 249-page manuscript, a 62-page discovery-paths document, and Lean 4 formalizations on GitHub under Apache 2.0 with zero unfinished proof goals.
  • Solved problems include the first explicit construction of non-sofic groups, a counterexample to Connes's rigidity conjecture, improved sphere packing bounds (first exponent improvement since 1978), and three Erdős problems closed.
  • Astra is a multi-agent model family designed for long-running tasks lasting hours or days; Sam Altman has showcased it in Washington, D.C., and it will require U.S. government approval before public release.
  • The Lean 4 certificates represent a structural shift from social validation — nine mathematicians signed off on OpenAI's May 2026 Erdős result — to machine-checkable proofs, a credibility improvement underscored by OpenAI's October 2025 false math claim that Thomas Bloom publicly debunked.
  • The broader math community has not yet peer-reviewed the proofs; the formalization manifest labels review status as 'agent-reviewed,' and the community must still confirm the formalized statements match the intended open problems.

Launch facts

Price:
Tokens for all ten solutions would have cost roughly $2,000 at Sol API rates
Availability:
Not yet publicly released; undergoing testing and planned U.S. government review

Open source

Repository:
OpenAI GitHub repository (Lean 4 formalizations)
License:
Apache 2.0
Language:
Lean 4

What happened

On August 1, 2026, OpenAI published a blog post titled "Ten advances in mathematics and theoretical computer science." The third paragraph contained the first official confirmation of a new model family: the results, OpenAI wrote, "were achieved by an internal version of Astra, our next major model."

The ten results span high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. Among the most notable: the first explicit construction of a non-sofic group (settling a question open since Mikhail Gromov introduced soficity in 1999), a counterexample to Connes's rigidity conjecture on von Neumann algebras, improved asymptotic upper bounds for high-dimensional sphere packing (the first improvement to the exponent since 1978), and resolutions of three problems from Paul Erdős's catalog, including problem 183 on multicolor Ramsey numbers.

OpenAI released a 249-page manuscript, a 62-page account of the model's discovery paths, and Lean 4 formalizations for each proof on GitHub under an Apache 2.0 license. The repository's "sorry" count — Lean's placeholder for unfinished proof goals — stands at zero, meaning every step across all ten formalized proofs is fully verified by the proof assistant's trusted kernel.

Astra is described as a multi-agent model family designed for long-running tasks lasting hours or even days. CEO Sam Altman has already showcased the technology to policymakers in Washington, D.C., and Astra is expected to be the first system evaluated under a planned U.S. government framework requiring federal pre-release approval. OpenAI has not determined whether it will launch publicly as GPT-6 or as a variant within the GPT-5 lineup. The total tokens used to find all ten solutions would have cost roughly $2,000 at Sol API rates.

Why it matters

The structural detail that separates this announcement from prior AI-math milestones is the Lean 4 formalization. When OpenAI's model disproved the Erdős unit-distance conjecture in May 2026, validation came from nine external mathematicians reading and signing off — a social signal that is strong but not independently reproducible without the same expertise. A Lean certificate is different: the proof either compiles or it doesn't, and anyone with a laptop can run the check.

This matters even more given OpenAI's track record. In October 2025, the company made a similar claim about AI-generated math solutions that turned out to be false. Thomas Bloom, the University of Manchester mathematician who curates the Erdős problems catalogue at erdosproblems.com, publicly dismantled that claim. The Lean certificates in the Astra release are a direct response to that credibility gap — a machine-checkable guarantee rather than a lab's assertion.

Still, caveats remain. OpenAI's formalization manifest labels the review status as "agent-reviewed," not peer-reviewed. A Lean-checked argument verifies that the formalized proof follows from its stated assumptions, but the broader mathematics community must still agree that the formalized statements correspond to the intended open problems. The closest-vector result strengthens worst-case hardness evidence relevant to lattice cryptography but does not provide an attack on deployed post-quantum systems.

The announcement also fits a broader trend of accelerating AI-led mathematical discoveries. In May 2026, an OpenAI model cracked a decades-old Erdős conjecture. Last month, Anthropic's Claude Fable 5 found a counterexample to the Jacobian conjecture, which had stood for nearly a century. New Scientist reports that hundreds of other AI-led discoveries have been made in recent months.

What to watch

  • Whether the broader mathematics community accepts the formalized Lean statements as faithful representations of the claimed open problems.
  • The U.S. government review process for Astra, which could become a template for future frontier model approvals.
  • Whether OpenAI releases Astra as GPT-6 or as a GPT-5 variant, and what pricing and access model it adopts.
  • Peer-review reception of the ten proofs and any subsequent capability disclosures.

What to do next

Developers

Clone OpenAI's GitHub repository and run the Lean 4 proof assistant on each certificate file to independently verify that all proofs compile with zero 'sorry' goals.

Lean certificates offer a concrete, machine-checkable way to assess whether the formalized statements match the intended mathematical claims, independent of trust in the lab. The repo is under Apache 2.0 license.

Founders

Assess whether long-running, multi-agent AI reasoning could be applied to your domain's hardest unsolved research problems.

Astra's demonstrated ability to tackle decade-old open problems for roughly $2,000 in token costs suggests potential for R&D-intensive industries beyond mathematics.

PMs

Track the U.S. government review process for Astra as a potential template for future frontier model approvals.

If regulatory approval becomes a standard gate before public release, product roadmaps involving frontier models will need to account for review timelines.

Investors

Monitor peer-review reception of Astra's ten proofs and any subsequent benchmark or capability disclosures.

Credible novel mathematical research would differentiate OpenAI's frontier capabilities, but validation by the broader math community is still pending and OpenAI's October 2025 false claim underscores the need for caution.

Operators

Evaluate whether multi-agent, long-running task architectures could improve internal research or analysis workflows.

Astra's design for hours-to-days problem-solving hints at a shift from single-turn interactions toward sustained autonomous reasoning.

How to test

  1. 1Clone OpenAI's GitHub repository containing the ten Lean 4 formalizations (released under Apache 2.0 license).
  2. 2Run the Lean proof assistant on each certificate file to independently verify that all proofs compile with zero 'sorry' goals.
  3. 3Compare the formalized Lean statements against the corresponding claims in the 249-page manuscript to confirm they match the intended mathematical results.
  4. 4Review the 62-page discovery paths document to understand Astra's reasoning process for each solution.

Caveats

  • Astra itself is not publicly available and cannot be tested directly; only the Lean 4 formalizations and supporting documents can be inspected.
  • The formalization manifest labels review status as 'agent-reviewed,' not peer-reviewed. The broader mathematics community has not yet independently validated the proofs.
  • A Lean-checked argument verifies that the formalized proof follows from its stated assumptions, but the community must still agree the formalized statement is the right one.
  • OpenAI's October 2025 false math claim, publicly debunked by Thomas Bloom, underscores why independent Lean verification matters — but also why community scrutiny of formalized statements remains essential.