Morning Edition · Tuesday, August 4, 2026Published at 2:16 AM EDT · New York
OpenAI Publishes Lean-Checked Proofs for Ten Open Mathematics Problems From an Unreleased Model
The company puts the compute cost at about $2,000 and released machine-checkable certificates. No result has been through refereed review, and OpenAI is judging the novelty of its own work.
OpenAI said on August 1 that an internal version of its next model family, Astra, produced new results on ten open problems in mathematics and theoretical computer science. It published a 249-page manuscript with certificates for every result, written in Lean 4, software that checks a proof step by step. The central claim is an explicit construction of a non-sofic group, a question open since the mathematician Mikhail Gromov introduced soficity in 1999. Other results cover sphere-packing bounds, coding theory, quantum complexity and lattice cryptography. OpenAI puts the compute bill at roughly $2,000.
The formalization matters because it removes the usual objection to machine-generated proofs. A Lean certificate can be checked by anyone with a laptop, and the identity of its author is irrelevant to that check. What Lean cannot check is whether the formal statement faithfully renders the problem mathematicians actually considered open, whether the definitions match the field's intent, or whether the historical framing is correct. Those are judgement calls OpenAI is currently making about its own output, and none of the ten results has cleared a journal.
Commentary around the release has gone well beyond the evidence. A widely shared post relayed a claim from Emad Mostaque, the former chief executive of Stability AI, that a one-billion-parameter model trained only on pre-1911 material rederived general relativity on its own. No paper, weights or reproduction accompany that claim, and it does not carry the evidentiary weight of the Lean artifacts. The same channel is circulating recursive self-improvement framing that the Astra release does not support. Ten formalized theorems is a research result, not evidence that a system is improving itself.
- If true, who benefits
OpenAI, which converts a $2,000 compute run into a capability claim that competitors must answer before Astra ships commercially, and the formal-methods ecosystem around Lean, whose certificates are becoming the settlement mechanism for disputed artificial intelligence claims.
- The nuance
The artifacts are real and checkable, with independent coverage confirming the 249-page manuscript and zero unproven steps in the Lean files, but OpenAI is grading the novelty of its own work and the precedent argues for caution, since Thomas Bloom, who maintains the Erdős problems database, called an earlier OpenAI claim that GPT-5 solved ten Erdős problems a dramatic distortion before calling the Astra results significant.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
The binding constraint on machine mathematics has moved from generating candidate proofs to certifying that the formal statement is the interesting one, which shifts scarce human effort from verification to problem curation. Formal-methods tooling and Lean-adjacent infrastructure gain, and so does any lab that can convert a $2,000 compute run into publishable research. The open question is whether independent mathematicians confirm that all ten statements were genuinely open, or whether several turn out to be known results in unfamiliar notation, which would make the announcement a strong systems demonstration rather than a research event.
What to watch
- Whether working group theorists publicly confirm that the non-sofic group construction resolves the question as the field understood it, the single check that decides how much of this claim survives.
- Whether OpenAI ships Astra to outside researchers rather than only publishing its outputs, since a model no one can run cannot be tested on problems OpenAI did not pick.
- Whether competing labs answer with their own Lean-certified results, which would turn formal verification into the default currency for capability claims in mathematics.
Observations to monitor, not financial advice.
Source: Polylog editors
Part of a tracked trend
AI Moves Into Autonomous Scientific Discovery and Clinical Care
Over the next 3-9 months, AI systems move beyond text tasks into running real scientific experiments and managing clinical care, backed by peer-reviewed and benchmarked evidence of chemist- and physician-level performance.
More from this edition
- Alibaba Ships Qwen3.8-Max at 2.4 Trillion Parameters and Promises the Weights Next Week
- OpenAI Publishes Internal Messages to Rebut Apple's Trade-Secret Suit
- Tencent's Hyra Agent Claims Record Results on 29 of 55 Open Mathematics Problems
- OpenAI Details the Full-Duplex Stack Behind GPT-Live, Built in Six Months
- An Open-Source Runtime Streams Mixture-of-Experts Weights From SSD to Run an 80-Billion-Parameter Qwen on a Mac
- Meta Sells Its Best Model by the Token After Years of Giving Weights Away
- New Papers Push Back on Paying Frontier Prices to Grade Model Output
- Meta's Segmentation and Vision Models Move Into Assistive Robotics and National Laboratory Science
- Researchers Propose an Executable Benchmark for the Decisions Agents Make Before They Answer
- Agent Tooling Turns Toward Traces and Verified Skill Claims
- OpenAI Publishes a Telco Deployment With Revenue Numbers Attached
Comments
0No comments yet.