OpenAI's Astra solves 10 long-standing maths problems
Can AI tackle frontier mathematics? OpenAI's Astra has solved 10 challenging problems, pushing the limits of AI reasoning.
A new AI milestone has arrived in mathematics. OpenAI has revealed that an internal version of its upcoming AI model, Astra, has achieved breakthroughs in 10 challenging problems spanning mathematics and theoretical computer science. Rather than solving routine equations, the model tackled long-standing open research questions that have challenged experts for years.
According to OpenAI, each result either solves a problem outright or makes substantial progress towards its solution. While the findings will now undergo scrutiny from the wider academic community, they prove how AI can contribute to original mathematical research.
Why mathematics is a true test for AI
Mathematics demands far more than producing the correct answer. Every conclusion must be backed by logical reasoning that can withstand detailed examination by experts. OpenAI says Astra generated complete mathematical arguments for all 10 research problems.
Researchers then worked with the model to prepare the findings as academic manuscripts. To strengthen confidence in the results, OpenAI also formalised every proof using Lean, an open-source proof assistant designed to verify mathematical logic.
Unlike traditional written proofs, which may contain hidden assumptions or overlooked gaps, Lean converts mathematical reasoning into a format that software can rigorously check. Although formal verification does not replace expert peer review, it provides an additional layer of confidence before the mathematical community evaluates the work.
Important note: Astra is an unreleased AI model developed by OpenAI.
Research areas where Astra made progress
The breakthroughs span several advanced branches of mathematics and theoretical computer science. Among the reported achievements are new results in high-dimensional sphere packing, which explores how efficiently spheres can be arranged in higher dimensions, and improved bounds for binary and spherical codes that are widely used in information theory.
OpenAI also reported progress on non-sofic groups, Connes's rigidity conjecture, the closest vector problem, quantum parallel repetition, arithmetic circuit complexity, Ehrhart's volume conjecture, multicolour Ramsey numbers and extremal graph theory.
Although these topics are highly specialised, they underpin important areas of computing, cryptography, optimisation and quantum information science.
A surprisingly accessible computing cost
One of the more unexpected details in OpenAI's announcement is the reported computational cost. The company estimates that generating the required solutions would require a number of tokens costing roughly $2,000 at Sol API pricing.
That figure does not reflect the entire research effort, including evaluation and verification. However, it suggests that advanced AI reasoning could become an increasingly practical research tool, enabling mathematicians to explore ideas, test hypotheses and investigate difficult problems more efficiently.
OpenAI’s AI Solves 80-Year-Old Geometry Puzzle, Mathematicians Verify Proof
A responsible approach to AI-generated research
OpenAI has also stressed the importance of transparency as AI becomes more involved in scientific discovery. The company argues that AI-generated proofs should be properly attributed and that presenting fully AI-produced work as entirely human-generated would misrepresent the research process.
The broader mathematics community will now evaluate Astra's results through peer review, independent verification and further analysis. That process will determine how significant each breakthrough ultimately proves to be.
Even so, the announcement marks an important landmark. AI is no longer limited to solving benchmark tests or assisting with calculations. It is beginning to contribute to original mathematical discovery, potentially reshaping how future research is conducted across mathematics, computer science and related scientific disciplines.