OpenAI and Anthropic AI Models Independently Refute the Exponential Time Hypothesis

On October 16, 2026, two competing artificial intelligence models independently solved one of the most stubborn puzzles in theoretical computer science. OpenAI and Anthropic have both published machine-generated proofs refuting the Exponential Time Hypothesis, a 25-year-old conjecture that defines the absolute speed limit for solving complex logic problems.[1][4][5]

The breakthrough centers on a classic logic puzzle known as 3-SAT, where a computer must determine if a vast array of true-or-false switches can satisfy a rigid set of rules simultaneously. For decades, computer scientists assumed that the computational work required to solve 3-SAT would always double with each additional switch, creating an unavoidable exponential explosion.[1][2]

That assumption, formalized in 1999 by computer scientists Russell Impagliazzo and Ramamohan Paturi, became the bedrock for hundreds of subsequent proofs across graph theory, databases, and computational biology. Researchers used the Exponential Time Hypothesis to prove that if 3-SAT was fundamentally slow, countless other problems in the field were equally intractable.[1][4]

Now, both OpenAI and Anthropic claim their internal research models have found a deterministic algorithm that beats that exponential curve. The AI-discovered method solves 3-SAT—and its harder counting variant, #3-SAT—in subexponential time, where the exponent's share shrinks as the number of variables grows.[4][5]

A simultaneous discovery

The timeline of the discovery has captivated the mathematics community. According to OpenAI, its unreleased model generated a natural language disproof of the hypothesis on October 16, which researchers formalized in the Lean 4 programming language the following day.[4][6]

Hours after OpenAI announced its findings on October 26, Anthropic published its own paper, titled "A 2^{o(n)} Time Algorithm for 3-SAT." Anthropic revealed that an internal Claude research model had independently found the core algorithm on the exact same day, October 16, using 14 million output tokens and zero human input.[5][6]

Both companies have released their machine-checked Lean proofs publicly on GitHub. The simultaneous, unsteered discovery by two separate architectures suggests that frontier AI models have crossed a threshold into autonomous mathematical reasoning, moving beyond pattern matching to generate novel, world-class theorems.[4][5][6]

The algorithm itself relies on a highly technical reorganization of the problem. By splitting the variables into groups that share only a few numbers, the AI's method recovers the exact count of satisfying assignments through rational interpolation, bypassing the need to check every combination one by one.[4][5]

Verification and the shepherd model

Because the proofs are dense and highly technical, verifying them requires immense human effort. Rather than simply dropping the results online, OpenAI has introduced a "shepherd" model, recruiting elite external academics to organize working groups that will digest and explain the AI's math.[3][4]

The computer science working group includes Ryan Williams of MIT, alongside Impagliazzo and Paturi—the very researchers who proposed the Exponential Time Hypothesis 25 years ago. While the academics stress that their participation is not an immediate certification of the proof, they are actively reviewing the Lean formalization.[1][4]

"Examining, questioning, explaining, and building on one another's work is central to how science advances," Paturi told The Washington Times. He noted that the three decades of research built on the hypothesis remain highly valuable for their structural insights, even if the foundational assumption falls.[1]

Anthropic's publication page makes a striking disclosure about the complexity of the machine's output. The company explicitly labels the paper as primarily AI-generated, noting that human understanding of the work currently covers only some parts of the formalized Lean 4 proof.[5]

What this means for computing

While the theoretical implications are seismic, the practical impact on everyday computing is zero. The AI-discovered algorithm is asymptotically faster, meaning it only beats brute-force guessing on problems with an astronomically large number of variables, estimated by researchers at more than 10 to the 108th power.[2][5]

The algorithm will not make commercial SAT solvers faster, nor does it prove that P equals NP—the ultimate prize in computer science. P versus NP asks whether hard problems can be solved quickly in polynomial time; the new algorithm lands 3-SAT in a middle zone, faster than exponential but still profoundly slow.[2][4][5]

However, the destruction of the Exponential Time Hypothesis forces a massive rewrite of theoretical computer science. A stronger cousin, the Strong Exponential Time Hypothesis, falls alongside it, reopening questions about the absolute speed limits of everyday tasks like comparing DNA sequences or measuring text differences.[4][5]

A broader scientific release

The ETH refutation was only one part of OpenAI's massive October 26 data dump. The company also claimed its models had resolved the Kakeya conjecture and Stein's restriction conjecture in all dimensions, utilizing a novel auxiliary partial differential equation discovered on October 18.[4][6]

More remarkably, OpenAI described a rigorous existence proof for pure Yang-Mills theory, a critical piece of the Yang-Mills mass gap problem that the Clay Mathematics Institute lists among its seven Millennium Prize Problems. A separate physics working group, led by theoretical physicist David Gross, is shepherding that specific result.[4][6]

Critics have expressed frustration with how the companies are managing access to the underlying models that produced these breakthroughs. Researchers argue that gatekeeping the AI systems prevents the broader academic community from interrogating how the models arrive at their insights or testing them on other open problems.[3]

For now, the mathematics and computer science communities are focused on the code in front of them. If the Lean proofs hold up to human scrutiny, October 16, 2026, will be recorded as the day artificial intelligence fundamentally altered the boundaries of human knowledge.[1][4][5]