OpenAI and Anthropic Claim Refutation of Exponential Time Hypothesis Pending Independent Verification

The refutation of a core 25-year-old computational conjecture challenges the theoretical hardness baselines that govern algorithm design and conditional complexity proofs across more than 800 published papers.

* The frontier lab announcements: On October 26, 2026, OpenAI published a preprint claiming an internal model derived a subexponential-time algorithm refuting the Exponential Time Hypothesis (ETH), formalized in the Lean interactive theorem prover. * The parallel Anthropic confirmation: Hours later, Anthropic announced an independent refutation derived via its Claude system, pointing to an identical breakdown of the conjecture's core premise. * The expert verification mobilization: OpenAI assembled an open review panel featuring theoretical computer scientists Ryan Williams and original ETH proposers Russell Impagliazzo and Ramamohan Paturi to independently audit the formal proofs without pre-endorsement.

* The bedrock of hardness theory: Formulated in 2001 by Russell Impagliazzo, Ramamohan Paturi, and Francis Zane, the Exponential Time Hypothesis posits that specific NP-complete problems—most notably 3-SAT—fundamentally demand exponential worst-case runtimes. * The structural dividing line: ETH does not resolve whether P equals NP; instead, it establishes a stricter conditional metric separating superpolynomial runtimes from strictly exponential runtimes, providing theoretical lower bounds for classical algorithms. * The downstream dependency footprint: More than 800 academic papers directly rely on ETH or its related conjectures to establish that faster exact algorithms for specific computational problems are impossible under standard assumptions.

The technical core of this event is not simply a theoretical challenge to a single conjecture; it represents an operational inflection point in how formal mathematical verification intersects with computational hardness. For nearly three decades, theoretical computer science operated under the working premise that solving worst-case 3-SAT instances required runtimes scaling as $2^{\Omega(n)}$. This boundary allowed computer scientists to construct clean conditional impossibility proofs across scheduling, routing, network design, and combinatorial optimization. When a problem was proven ETH-hard, teams ceased looking for faster exact algorithms.

If the internal models at OpenAI and Anthropic have produced verified algorithms running in subexponential time (such as $2^{o(n)}$), that lower bound evaporates. The mechanism behind both labs' claims relies heavily on interactive theorem proving—specifically formalizations in languages like Lean—which allow logical statements to be verified by proof-checking kernels rather than human intuition alone. By targeting 3-SAT with subexponential routines, the models suggest that worst-case instances of canonical NP-complete problems are structurally softer than the academic consensus assumed.

The strategic incentive driving both labs is clear: the transition from statistical pattern matching to formal mathematical discovery. Showing that frontier AI architectures can overturn long-standing human conjectures by constructing machine-checkable algorithmic frameworks signals an expansion into formal reasoning. At the same time, Anthropic issuing a parallel claim within hours illustrates intense competitive alignment in automated formal science.

Yet, structural skepticism remains necessary. As researchers including Timothy Gowers, Gary Marcus, and John D. Cook have noted, a refutation of ETH does not mean P equals NP. It leaves polynomial-time tractability unresolved while reopening the exact boundaries of non-polynomial runtimes. If the proofs survive peer review, theoretical computer science must reorganize 25 years of algorithmic architecture, revisiting conditional boundaries on hundreds of fundamental combinatorial problems.

* Theoretical computer science faculty: Researchers managing research portfolios tied to fine-grained complexity must re-evaluate the premises of over 800 conditional hardness proofs. * AI formal reasoning teams: Frontier AI research divisions validating automated formal theorem provers gain immediate incentive to apply verification agents to other open mathematical conjectures. * Enterprise algorithm developers: Teams designing exact solvers for high-stakes optimization, logistics, and scheduling must monitor whether subexponential theoretical bounds yield practical algorithmic implementations. * Global tech hubs and research labs: Emerging innovation ecosystems evaluating intellectual capital distribution will see renewed institutional priority placed on formal logic and computational complexity talent.

* The Impagliazzo-Paturi-Williams audit findings: Formal review findings from the conjecture’s original authors will determine whether the subexponential algorithm contains logical flaws or fully withstands academic scrutiny. * Lean verification kernel confirmations: Independent execution of the formalized Lean codebases by external research groups will show whether the computational proofs are machine-valid without non-standard axioms. * Algorithmic translation into executable code: Watch for whether the theoretical subexponential bounds translate into practical algorithmic speedups for benchmark NP-complete problem instances.