AI Models from Frontier Labs Challenge Core Computational Complexity Assumption

Image: TechCrunch · Source

OpenAI and Anthropic AI models claim algorithms that refute the longstanding Exponential Time Hypothesis, a foundational assumption about computational difficulty. This could reshape understanding of the P vs NP problem and impact cryptography and computer science.

On October 27, 2026, OpenAI and Anthropic announced groundbreaking claims that their AI-developed algorithms refute the Exponential Time Hypothesis (ETH), a cornerstone assumption in computational complexity theory. ETH posits that certain problems, notably 3-SAT—a type of Boolean satisfiability problem—cannot be solved faster than exponential time as their size grows. This hypothesis underlies much of theoretical computer science and informs expectations about problem hardness and cryptographic security.

OpenAI released proofs developed by their models on GitHub, sparking rapid and mixed reactions from the scientific community. To coordinate understanding and validation, OpenAI engaged leading computer scientists, including MIT Professor Ryan Williams and UC San Diego’s Russell Impagliazzo and Ramamohan Paturi, who originally formulated ETH in 1999. These experts are organizing working groups and encouraging open feedback to rigorously assess the findings.

According to the published research, an AI model named Claude (from Anthropic) discovered deterministic algorithms solving the 3-SAT problem and even its counting variant (#3-SAT) in sub-exponential time, specifically in roughly 2^{O(n f n / sqrt{log n})} time. This asymptotic improvement suggests ETH is false, challenging previous limitations on algorithmic efficiency.

The Exponential Time Hypothesis is closely tied to the P vs NP question, which asks if problems whose solutions can be quickly checked (NP) can also be quickly solved (P). While ETH provides strong evidence that many NP problems are inherently hard, the new AI-derived algorithms imply some of these problems may be tractable faster than expected, potentially nudging scientific opinions about the plausibility of P = NP.

However, Ryan Williams emphasizes caution, noting that the AI-generated algorithms might not represent optimal solutions and the field must collectively explore the implications. Many prior research results built on ETH may need reevaluation if these claims hold under peer scrutiny.

Such a paradigm shift carries important cybersecurity ramifications. Cryptographic protocols often rely on the hardness of NP problems and ETH; improvements in solving those problems might undermine certain cryptosystems. Williams confirmed OpenAI checked their algorithms against cryptographic security and found no immediate vulnerabilities.

This evolving collaboration between AI labs and academia illustrates a new model of scientific discovery. Despite concerns about integrating AI contributions, experts remain optimistic that such advancements will push researchers to explore more ambitious questions and adapt foundational understanding in theoretical computer science.

The AI-developed algorithms and formal proofs are publicly accessible via Anthropic’s GitHub, including Lean 4 formalizations that enable community verification. OpenAI and Anthropic plan to fund conferences and working groups to deepen discourse.

The significance of these developments lies not only in potential breakthroughs solving classic computer science problems but also in their demonstration of AI as an active partner in advancing scientific frontiers.