Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, targeting the hardest technical challenges with gains in coding, reasoning and vision. Fable 5.2 scored 41.3% on the Remote Labor Index, roughly double Opus 5.5 and GPT-6 Astra, and 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours versus 11 days for Anthropic's first formalization. Mythos 5.2 also cut sandbox-escape rates to 3.1% from 39% in Mythos 5.
Formal Proof
形式化证明
The process of expressing and verifying mathematical theorems using a computer-verifiable logical language
Curated Intelligence
Anthropic has upgraded Claude Fable and Claude Mythos to version 5.2, with gains in coding, reasoning and vision. Both share the same underlying model: Fable ships with cybersecurity and biology safeguards for general availability, while Mythos is gated behind verification programs for security and life-sciences organizations. In prerelease testing the models completed otherwise-infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in black hole physics, and formalized Fermat's Last Theorem in 30 hours.
TScert is a sound type checker for TypeScript that accepts a function only when accompanied by a proof, validated by a checker whose soundness is itself proved in Lean. It does not trust types inferred by TypeScript or related tools, addressing tsc's unsoundness around any, type assertions, unchecked index reads and covariant arrays.
Sam Altman addressed mathematicians' backlash against OpenAI's research publications, admitting flaws in rigor and context while promising better integration with the academic community. Despite protests from Fields Medalists and controversy over withholding Yang-Mills theory details, he affirmed that humanity deserves to understand fundamental laws and will not halt scientific progress. OpenAI also pledged $5 million to the Simons Foundation to help mathematicians adapt to AI-generated discoveries.
Startup Physical Superintelligence (PSI) announced that its proprietary AI research system has made breakthroughs on five long-standing open problems in quantum information theory, covering core areas such as entanglement distillation and quantum channel capacity. The associated papers, Lean formal proofs, and verification code have been published on GitHub for community review, marking a new phase in the application of AI to fundamental physical science research.
Axiom Math announced that a cluster of dozens of AI agents completed the formalization of the Classification of Finite Simple Groups (CFSG) in Lean within just 24 days. The project generated approximately 43.1 million lines of code and 2.05 million declarations, achieving what took human mathematicians decades, while also uncovering a nearly 40-year-old proof gap and constructing new group theory results.
OpenAI announced major breakthroughs in mathematics, computer science, and physics, including claims of proving the Kakeya conjecture and disproving the Exponential Time Hypothesis. While the company introduced expert 'guides' to manage publication, researchers criticized the lack of human contribution and the difficulty of verifying AI-generated formal proofs.
In 2026, large models achieved remarkable progress in mathematics. OpenAI's internal model disproved the Erdős unit distance conjecture and solved long-standing problems like the Cycle Double Cover and Navier-Stokes equations. Meanwhile, Anthropic formalized Fermat's Last Theorem and advanced bounds on Riemann's hypothesis, signaling a paradigm shift from computational aid to autonomous mathematical discovery.
The mathematics community is distancing itself from OpenAI due to perceived lack of integrity and exaggerated marketing, favoring open-source tools like the Lean prover for their honest and impactful contributions. This shift highlights a broader industry trend towards transparency and verifiable results over hype.
While recent AI models excel at theorem proving, sparking calls to formalize all mathematics, senior mathematicians warn against rushing. They argue that prioritizing speed over quality leads to low-value code where cleanup costs exceed rewriting, potentially causing lasting damage to mathematical knowledge bases. The authors advocate for a cautious, high-quality approach rather than an aggressive mass-formalization effort.