OpenAI has unveiled new research results from Astra, an unreleased artificial intelligence (AI) model, claiming it has solved 10 mathematics problems that had remained unsolved for at least a decade.
The breakthroughs span multiple branches of mathematics and theoretical computer science, including cryptography, quantum computing, geometry, graph theory, and computational complexity.
According to OpenAI, Astra independently generated the mathematical arguments, which were later refined into research papers by human researchers with assistance from the same AI model.
The proofs were then formally verified using Lean, an open-source proof assistant designed to validate logical arguments mathematically.
To provide greater transparency, OpenAI also published Astra's reasoning narratives alongside each proof, allowing researchers to examine how the AI approached the problems.
Key Mathematical Breakthroughs
OpenAI said Astra produced new results in several major research areas, including:
High-dimensional sphere packing
Binary and spherical error-correcting codes
Arithmetic circuit complexity
The Closest Vector Problem (CVP)
Quantum parallel repetition
Ehrhart's volume conjecture
The company also claimed the model:
Established the existence of non-sofic groups
Produced a disproof of Connes' Rigidity Conjecture
Solved three open problems proposed by renowned mathematician Paul Erdős in extremal graph theory and Ramsey theory
If independently verified, these results could represent significant advances across multiple mathematical disciplines.
AI-Generated Research Papers
OpenAI stated that generating the mathematical solutions using Astra would have cost approximately $2,000 through its Sol API pricing.
Researchers then used the same model to prepare the accompanying manuscripts before the proofs were formally verified in Lean.
The company emphasized that human researchers remained involved throughout the process, particularly in preparing the papers and ensuring the proofs met formal verification standards.
Follows Earlier Mathematical Discovery
The announcement comes months after OpenAI revealed that another unreleased AI model had generated a claimed disproof of the famous Erdős Unit Distance Conjecture during internal testing. Astra represents the company's latest effort to demonstrate AI's potential in advanced mathematical research.
ChatGPT for Academic Researchers Launched
Alongside Astra's announcement, OpenAI introduced ChatGPT for Academic Researchers, a new initiative aimed at supporting scientific research.
Under the program:
100,000 scientists and mathematicians will receive free access to OpenAI's latest ChatGPT models.
The initiative is designed to accelerate research, collaboration, and problem-solving across scientific disciplines.
Experts Urge Caution
Despite the impressive claims, researchers cautioned against viewing Astra as evidence that AI's broader reliability challenges have been solved.
Cognitive scientist and AI researcher Gary Marcus argued that strong performance in mathematics does not necessarily translate into overall reliability.
In a post on X, he noted that Astra may excel in certain mathematical domains but could still suffer from issues common to generative AI systems, including hallucinations, difficulty interpreting documents accurately, and inconsistent rule-following.
OpenAI has not yet released Astra publicly. The company said the mathematical results will now be evaluated through peer review and independent verification by the wider mathematics community.
Whether Astra's claimed breakthroughs withstand academic scrutiny will determine their long-term significance, but the announcement marks another major milestone in the growing role of artificial intelligence in advanced scientific research.
