Google's AlphaProof Nexus Solves 9 Unsolved Math Problems for Dollars

Google DeepMind’s AlphaProof Nexus, an AI system pairing large language models with formal proof-checking, has solved nine out of 353 open Erdős problems and 44 of 492 open conjectures from the Online Encyclopedia of Integer Sequences. The breakthrough, which cost only a few hundred dollars per problem, demonstrates a new frontier in AI-driven formal verification that could change how critical software is built.

"Organizations should use caution with vibe coding without verification, as AI systems are rapidly moving into environments where correctness is no longer optional," Eve Bodina, founder and CEO of Logical Intelligence, a rival AI lab, said in a recent statement. "Formal reasoning benchmarks are increasingly important because they force AI systems to operate in environments where correctness is mathematically enforced."

The results were documented in an arXiv preprint (2605.22763v1) published on May 21, 2026. AlphaProof Nexus works by generating a mathematical proof with a large language model and then using the Lean proof assistant to check every logical step for correctness. This "agentic loop" iterates on proposed proofs until they are formally verified, a direct response to the persistent problem of AI hallucination that has plagued enterprise adoption.

This development moves AI from generating plausible-sounding text to producing provably correct logic. The implications extend far beyond academia, threatening to change the economics of smart contract auditing, cryptographic protocol design, and zero-knowledge proof generation—fields where a single logical error can lead to catastrophic financial losses.

A New Front in the AI Verification Arms Race

Google is not alone in using AI to tackle frontier mathematics. OpenAI recently announced one of its general-purpose models disproved a central conjecture related to the Erdős planar unit distance problem by finding a novel counterexample. While DeepMind’s AlphaProof Nexus proved decades-old conjectures correct, OpenAI’s model found a flaw in a long-held mathematical belief. Both achievements, however, relied on elite human mathematicians to check, refine, and interpret the AI’s output, pointing toward a new division of labor between human and machine.

The differing approaches highlight a key trend: the AI industry is moving beyond benchmark scores and toward solving open problems where the answers are not known. This pivot from curated tests to frontier research is a critical step in demonstrating AI’s value as a collaborator in science and engineering, not just a tool for summarization. The core challenge remains trust, as AI-generated hallucinations continue to appear in courtrooms and academic papers.

From Academic Puzzles to Commercial Reality

The race to commercialize this technology is already underway. Logical Intelligence, an AI lab focused on energy-based reasoning models, recently announced its agent, Aleph, has solved 99.4% of the PutnamBench, a benchmark for advanced mathematical theorem proving. This performance significantly outperforms systems from ByteDance and other competitors.

Logical Intelligence is already deploying Aleph in production verification workflows, including work with the Ethereum Foundation’s cryptographic libraries. This transition from academic proof-of-concept to production-grade verification for critical infrastructure shows that a new market is emerging. Companies are building AI not just to generate code, but to prove it is correct before it ever reaches a production environment where failures have real-world consequences.

For investors, the key insight is that the ability to generate provably correct output is a foundational requirement for scaling AI in mission-critical systems. This shift directly addresses the primary weakness of current generative models: their tendency to confabulate under pressure. While Alphabet's (GOOGL) achievement with AlphaProof Nexus reinforces its leadership in AI research, the emergence of specialized firms like Logical Intelligence indicates a new infrastructure layer for "verified AI" is being built. This technology will be essential for any industry, from finance to energy, that cannot afford to be wrong.

This article is for informational purposes only and does not constitute investment advice.