OpenAI’s October 26, 2026, announcement has drawn attention to a major claim in theoretical computer science: an artificial intelligence (AI) system may have refuted the Exponential Time Hypothesis (ETH). The result could reshape how researchers study difficult computing problems, but independent experts are still examining the proof and its implications.
According to OpenAI’s October 26 announcement, its internal AI system generated a natural-language argument against the Exponential Time Hypothesis on October 16, 2026, and the result was subsequently formalized in Lean on October 17. The company made the written proof and supporting mathematical files available so researchers could examine the findings.
The Exponential Time Hypothesis is a long-standing idea about a computing problem called 3-SAT. In simple terms, 3-SAT asks whether a set of true-or-false choices can satisfy all the rules in a logical formula. ETH says there is no algorithm that can solve every 3-SAT problem in subexponential time as the number of variables increases. Researchers have used this assumption for decades to understand the difficulty of many other computing problems.
OpenAI says the AI-generated result refutes ETH and includes a computer-checked Lean formalization. The project’s published scope of the Lean formalization describes a subexponential algorithm for certain Boolean formulas. Although this is a potentially important development, experts still need to examine whether the formalized mathematical statement and its interpretation support the full claim about ETH.
Separately, Anthropic also published an AI-generated 3-SAT paper and Lean proof in October 2026, reporting another result that challenges the Exponential Time Hypothesis. The research was produced using an internal Claude model and describes a theoretical algorithm for solving 3-SAT problems. Anthropic also explains that the finding does not demonstrate faster performance for practical computing problems today.
If the findings hold up under expert review, they could change how scientists approach some of the hardest problems in computer science.
ETH has been used to study difficult computing problems in areas such as database management, graph theory, computational biology, and algorithm design. If the new result survives expert review, researchers may need to reconsider some earlier conclusions about how difficult these problems are. However, challenging a theoretical assumption does not automatically make these tasks easier or faster in real-world applications.
OpenAI’s latest announcement follows earlier AI research on major mathematical problems. TechBonafide previously covered OpenAI’s AI-generated Navier–Stokes proof, another example of artificial intelligence entering advanced scientific research while raising questions about verification and expert review. These developments highlight how AI systems are increasingly being used to investigate problems that have challenged researchers for decades.
Earlier in October 2026, OpenAI released hundreds of AI-produced mathematical papers covering difficult research problems, including topics connected to theoretical computer science. TechBonafide explained this development in its report on 722 AI-generated mathematics manuscripts. The latest ETH announcement represents another important claim about how advanced AI systems could contribute to complex mathematical research.
OpenAI says a group of computer science experts led by Russell Impagliazzo and Ramamohan Paturi of the University of California, San Diego, along with Ryan Williams of MIT, will examine and explain the ETH result. Their work is intended to help researchers understand the proof, evaluate its meaning, and identify its wider consequences. The involvement of these experts should not be confused with a completed independent review or universal acceptance of the claim.
Despite these concerns, AI is increasingly becoming a powerful research assistant. Human experts remain important for selecting meaningful problems, checking results, explaining discoveries, and finding practical applications.
For everyday users, the claim does not mean that phones, laptops, or apps will suddenly become faster. A subexponential algorithm is not the same as a polynomial-time algorithm, so the result would not by itself prove that P equals NP or that widely used encryption has been broken. Turning this type of mathematical research into practical technology improvements would require additional studies, testing, and software development.
As researchers continue reviewing these results, AI could play an increasingly important role in shaping future computing technology, improving problem-solving methods, and expanding scientific knowledge.
