OpenAI announced on October 26, 2026 that one of its AI models produced a proof that the Exponential Time Hypothesis (ETH), a conjecture that computer scientists have treated as a bedrock assumption for about thirty years, is false. The proof was checked in Lean, an interactive theorem prover that verifies each logical step by machine, and OpenAI handed the paper to the complexity theorists who originally formulated the hypothesis to review and shepherd it. The result does not show that P equals NP, and OpenAI President Greg Brockman said as much the same day. What it does show is that a class of problems long assumed to require brute-force search admits a sub-exponential shortcut, and that shifts the ground under a large body of theoretical work.
What the Exponential Time Hypothesis claimed
ETH concerns Boolean satisfiability, often shortened to SAT: given a logical formula built from variables that can each be true or false, is there an assignment that makes the whole formula true? The hypothesis, proposed roughly three decades ago by Russell Impagliazzo, Ramamohan Paturi and Francis Zane, asserted that solving the k-CNF version of this problem requires time that grows strictly exponentially with the number of variables. In plain terms, it said there is no meaningfully better strategy than trying configurations until one works.
That assumption became a workhorse. Hundreds of papers in graph theory, computational biology, database theory and computational geometry used ETH as a premise to derive conditional hardness results, statements of the form "if ETH holds, then this other problem cannot be solved quickly either." Removing the premise does not make each of those downstream claims wrong, but it does mean they no longer stand on the foundation their authors assumed. A useful way to picture it: an entire district of theoretical structures was built on one plot of land, and the land turned out not to exist in the form everyone believed.
What the new paper shows
OpenAI's paper, titled "The Exponential Time Hypothesis is False," carries an original date of October 17, 2026 and a revision date of October 25, 2026. According to the accompanying announcement, the model found the disproof on October 16 and the Lean formalization was completed on October 17. The central theorem presents a deterministic algorithm that solves k-CNF satisfiability in time 2 to the little-o of n, the formal notation for sub-exponential growth. The paper pairs the machine-checked Lean proof with a natural-language explanation.
Two points of framing are worth keeping separate:
| Statement | Status after the announcement |
|---|---|
| ETH is true | Refuted, per OpenAI's Lean-verified proof |
| SAT can be solved in sub-exponential time | Established by the paper's algorithm |
| P equals NP | Not shown; Brockman says it still probably does not |
| Modern encryption is broken | Not shown; depends on P vs NP, which remains open |
Anthropic also appears to have arrived at a very similar result at almost the same time, which, if confirmed, would mark a notable convergence between two independent AI systems.

▲ Exponential search versus fast verification
Why "sub-exponential" is not the same as "fast"
The distinction between P, NP and exponential time can be understood with a few everyday objects.
A combination padlock with one dial of ten digits takes at most 10 attempts to open by brute force. Two dials raise that to 100, three to 1,000, and ten dials to 10 billion. Each added dial multiplies the search space, which is exponential growth: small increases in size produce explosive increases in work.
Sudoku illustrates the other half of the picture. Filling an empty 9-by-9 grid by hand is slow, but checking a completed grid takes seconds, because you only need to confirm that each row, column and 3-by-3 block contains the digits one through nine. Problems whose proposed answers can be checked quickly belong to the class NP, which stands for nondeterministic polynomial time.
Class P, polynomial time, covers problems that get harder in a predictable, manageable way as they grow. Sorting one basket of laundry might take ten minutes; sorting five baskets takes proportionally longer, not billions of years.
The P versus NP question asks whether every problem that is easy to check is also easy to solve. For more than fifty years, researchers have tried to either find a polynomial-time algorithm for an NP-complete problem or prove that none exists. Neither happened, and the consensus drifted toward assuming P does not equal NP.
ETH was a sharper statement than P not equal to NP. It said not only that SAT lacks a polynomial-time algorithm, but that it needs fully exponential time. The new algorithm breaks that sharper claim. Rather than blindly spinning every dial, it behaves more like a lockpicker who listens to the tumblers: establishing that one dial sits at nine can logically rule out thousands of combinations elsewhere, so whole branches of the search are never visited. The running time still grows faster than any polynomial, as far as anyone has shown, but slower than a pure exponential.
A mountain analogy makes the hierarchy concrete. Getting from point A to point B across a mountain by climbing over the summit is exhaustive search, the world where P does not equal NP. A flat tunnel at ground level would be P equals NP. The ETH refutation does not open a ground-level tunnel, but it reveals tunnels bored through the mountain at intermediate heights, passages that decades of human effort had not located.
Why the result matters for cryptography, and why it does not break it
Essentially all modern cryptography depends on the assumption that recovering a secret key is an intractable NP search problem, a point made in Scott Aaronson's survey paper "P =? NP." Passwords and encryption keys are designed to take millennia to guess and milliseconds to verify. That asymmetry is the whole point, and it traces back to the earliest thinking in the field: a 1955 letter from John Nash to the National Security Agency argued that breaking a cipher requires searching a space that grows exponentially with key length, and a 1956 letter from Kurt Gödel to John von Neumann framed the metamathematical core of the same question.
If P were ever shown to equal NP with a practical algorithm, internet security, electronic banking and cryptocurrency systems would be fundamentally compromised and digital infrastructure would need an overhaul. The same shortcut would make enormous optimization problems tractable: shift scheduling, supply chain routing, semiconductor chip layout, protein structure modeling and drug design would all accelerate dramatically. An OpenAI researcher has also noted that foundational breakthroughs in physics, mathematics and computer science would act as a tailwind for AI progress itself.

▲ Cryptography and optimization at stake in P vs NP
None of that follows from the ETH disproof alone. Brockman's post that "p still probably doesn't equal np" drew more than 350,000 views, and other commenters emphasized that the result points toward sub-exponential algorithms rather than a polynomial-time collapse. There is a second caveat as well: even a polynomial-time algorithm can be useless in practice if its constant factors are astronomical. An algorithm that only outpaces brute force once the problem exceeds thousands of years of manual attempts offers little to anyone. One commentator on X took that idea further and argued that if P equals NP but only with constants inaccessible to inhabitants of our world, the most reasonable conclusion would be that reality is an engineered simulation with deliberate speed limits. That is a thought experiment, not a finding, but it captures why the gap between theoretical and practical efficiency matters so much here.
How OpenAI handled the rollout
OpenAI is taking no formal authorship credit. Instead, it arranged for Impagliazzo and Paturi of UC San Diego, along with Ryan Williams of MIT, who had earlier proven refutations of stronger versions of ETH, to review and shepherd the paper. The approach appears designed to give the result the scrutiny of the people best positioned to find flaws and to avoid the kind of backlash that followed earlier AI announcements.
The ETH paper is one of three results in the same announcement. OpenAI also published a rigorous existence proof for pure Yang-Mills theory, built on Tadeusz Bałaban's approach to constructive quantum field theory and formalized in Lean on October 16, 2026. That addresses a major component of the Clay Mathematics Institute's Yang-Mills and Mass Gap problem but is not a full solution to the Millennium Prize problem as officially stated. A third result concerns local smoothing and the Kakeya conjecture in harmonic analysis, a problem posed by Christopher Sogge in 1991. About thirty days earlier, OpenAI had released an analytical proof and Lean formalization for Navier-Stokes existence and smoothness, which is still undergoing academic peer review. Taken together, AI systems have produced proofs for multiple major conjectures within roughly a 90-day span.
Scott Aaronson, speaking on a podcast about 18 hours before the announcement, observed that theoretical physicists who once assumed AI would not touch their field are now watching its progress nervously, and he has argued that mathematicians are not the arbiters of who gets to work on their problems.
A historical pattern worth remembering
The result fits a pattern that Emanuele Viola, a computer science professor at Northeastern University, laid out in a 2018 essay titled "I Believe P=NP." In the 1950s, Andrey Kolmogorov conjectured that multiplying two n-bit integers required quadratic time; his 23-year-old student Anatoly Karatsuba found a faster algorithm within a week. In 1968, Volker Strassen overturned the assumption that standard matrix multiplication was optimal. Viola's thesis is that researchers repeatedly declare computational barriers that clever algorithms later surpass, and that humanity has underestimated the reach of efficient computation. The ETH disproof appears to be another entry in that list, though it remains unclear whether it represents a genuine practical shortcut or a modest push on a theoretical boundary.
What to take from this
- The Exponential Time Hypothesis, assumed for about thirty years, has been refuted by an AI-generated proof that was machine-checked in Lean and is being shepherded by the hypothesis's original authors.
- The proof gives a deterministic sub-exponential algorithm for k-CNF satisfiability. It does not show P equals NP, and Greg Brockman has said P still probably does not equal NP.
- Encryption, banking and cryptocurrency security are not shown to be broken by this result.
- Translating the theoretical bound into real software speedups will take substantial further research, and constant factors may limit practical use.
For readers who follow this space, a few practical steps follow. Researchers whose results depend on ETH as a conditional hardness assumption should re-examine those results. Anyone evaluating future AI-generated mathematics should look for formal verification in a proof assistant like Lean before treating a claim as settled, and should check whether a claimed algorithm has feasible constants and exponents rather than assuming it is deployable. And for everyday security, the lock analogy still holds: each added character multiplies an attacker's search space, so password length remains the simplest defense available.


