OpenAI AI Model Cracks Major Math and Physics Proofs

Artificial intelligence has crossed a new threshold that few predicted would arrive this decade. On October 26, 2026, OpenAI announced that one of its internal AI models has produced three major scientific proofs — including a disproof of a widely believed conjecture in computational complexity theory and a result touching one of the Clay Mathematics Institute’s Millennium Prize problems in physics. The claims, formalized in the Lean proof assistant before public release, are now being handed to academic reviewers for independent verification.

Three Results That Could Reshape Multiple Fields

The announcement, detailed in a blog post titled “Building Scientific Understanding with Academic Communities,” describes three distinct results generated by an unreleased internal model. Each addresses a long-standing open problem in mathematics, computer science, or theoretical physics.

• Exponential Time Hypothesis disproof: The model found a disproof of the Exponential Time Hypothesis (ETH) on October 16, 2026, and formalized it in Lean the following day. ETH has been widely believed to be true since it was proposed in the late 1990s by Russell Impagliazzo and Ramamohan Paturi. If the disproof holds, it collapses hundreds of conditional hardness results that have anchored algorithm complexity theory for nearly three decades.
• Kakeya and local smoothing conjectures: The model produced a proof of Christopher Sogge’s 1991 local smoothing conjecture in all dimensions. This result would also settle the Kakeya conjecture and Stein’s restriction conjecture — two of the most important open problems in harmonic analysis. OpenAI generated this proof on October 18 and formalized it in Lean the next day.
• Yang-Mills existence: The model produced a proof that pure Yang-Mills theory exists in four dimensions, which is part of one of the Clay Mathematics Institute’s seven Millennium Prize problems. However, the result does not address the mass gap component of the prize, so it does not fully satisfy the Clay conditions.

A New Verification Model

What makes this announcement unusual is not just the mathematical claims themselves, but the process OpenAI has chosen for verification. Rather than submitting papers to traditional academic journals, the company is routing each result through named external shepherds who will lead independent scrutiny over the next 90 days.

The computer science claim has been handed to Russell Impagliazzo and Ramamohan Paturi at UC San Diego — the original proposers of ETH — along with Ryan Williams at MIT. In an open letter, the three wrote that their involvement should not be taken as certification of the claim. “A result of this significance requires careful, independent scrutiny,” they stated, adding that if the claim is correct, “it will change an important part of theoretical computer science.” Williams confirmed the group signed no non-disclosure agreements and will not be paid.

The physics result is being handled differently. A group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov, has chosen not to release the AI-produced material publicly for now. Instead, they will write up the work themselves before any public artifacts appear. OpenAI has published the ETH disproof and math proofs on GitHub as Lean files alongside model write-ups, but the Yang-Mills artifacts remain withheld pending the panel’s review.

Why the ETH Disproof Matters

The Exponential Time Hypothesis, proposed in 1999, essentially states that certain computational problems cannot be solved significantly faster than exhaustive search. It has served as a foundational assumption in computational complexity theory for over 25 years. Hundreds of research papers have built conditional results on top of ETH — showing that if ETH is true, then various other problems are also hard.

If the disproof survives peer review, the consequences ripple across theoretical computer science. Many of those conditional hardness results would no longer hold, and researchers would need to revisit assumptions about what can and cannot be computed efficiently. The result also has implications for the famous P versus NP question, as ETH was seen as a stepping stone toward understanding the boundary between tractable and intractable problems.

Formal Verification as the New Standard

One notable aspect of this announcement is OpenAI’s reliance on the Lean proof assistant. Lean is a formal verification tool that can mechanically check mathematical proofs, ensuring each logical step is valid. By formalizing all three results in Lean before making any public claims, OpenAI is establishing a new standard for AI-generated mathematical work.

This approach addresses a legitimate concern in the academic community: AI models can produce convincing-looking proofs that contain subtle errors. Machine-checked formalization does not guarantee the result is correct in the way mathematicians care about — the model could still be proving a different statement than intended — but it does eliminate an entire class of errors related to logical gaps.

OpenAI also stated that it does not seek academic credit for the three results and has given $5 million to the Simons Foundation for workshops and conferences on AI-produced mathematical work. This positioning suggests the company views these results as demonstrations of capability rather than traditional academic contributions.

Context Within a Broader AI Moment

The announcement arrives during an extraordinary week for the AI industry. On the same day, Anthropic filed its S-1 registration with the SEC for a Nasdaq IPO under the ticker ANT, reporting $16.3 billion in first-half 2026 revenue — a 1,300 percent increase year over year. The company posted its first operating profit of $974 million in the second quarter. Anthropic’s IPO could value the company between $1.5 trillion and $2 trillion, potentially matching or exceeding the amount SpaceX raised in its own IPO.

Meanwhile, Moonshot released Kimi K3.1 Pro, which Artificial Analysis ranked fourth on its Intelligence Index with a score of 55. The Chinese company plans to publish the model’s weights openly by November 4. South Korean police also arrested seven suspects in connection with AI-assisted cyber intrusions on roughly 90 organizations between September 2025 and October 2026.

These parallel developments paint a picture of an industry that is simultaneously advancing the frontiers of scientific knowledge, entering public markets at historic valuations, and creating new security challenges that law enforcement is still learning to address.

What Happens Next

The 90-day verification period will be critical. For the ETH disproof, Impagliazzo, Paturi, and Williams will examine whether the model’s argument holds up under rigorous scrutiny. If it does, theoretical computer scientists will need to reexamine decades of conditional results. For the Kakeya and local smoothing proofs, the mathematical community will need to verify that the formalized versions match the intended mathematical statements.

The Yang-Mills result faces the longest road to public acceptance. The Nobel-led panel’s decision to write up the work independently means the result will not be available for community review until that process is complete. And even if the existence proof is confirmed, the mass gap — the requirement that Yang-Mills theory predicts a positive lower bound on particle masses — remains open.

Regardless of the verification outcomes, the announcement marks a shift in how AI companies present scientific claims. By combining formal verification with transparent external review processes, OpenAI is establishing a template that could become standard for future AI-generated scientific results. The next 90 days will determine whether that template produces results the scientific community can trust.

---

Edited by Palawan @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous