OpenAI says unreleased Astra model made 10 major advances in mathematics
New Delhi: OpenAI says an internal version of its unreleased Astra model has produced ten new results across mathematics and theoretical computer science. The work covers areas such as geometry, coding theory, group theory, quantum complexity and cryptography.
The announcement, published on August 1, 2026, is a major scientific claim and will now face close examination by mathematicians. OpenAI describes the collection as a mix of solved problems and substantial progress on long-standing questions, rather than ten identical breakthroughs. The company has released a 249-page technical manuscript containing the arguments.
How Astra produced the mathematical results
According to OpenAI, Astra generated the main mathematical arguments during internal testing. Human researchers then prepared the work as manuscripts with help from the same model.
The model later formalised each argument using Lean, a proof assistant that lets computers check logical steps against strict mathematical rules. This provides another layer of checking, but it does not replace outside review by specialists.
OpenAI estimates that generating the successful results used enough tokens to cost around $2,000, or nearly ₹1.90 lakh, at its current Sol API rates. The figure appears to cover the total search for the ten selected results, rather than ₹1.90 lakh for every individual solution.
What problems did OpenAI Astra work on?
The ten results include:
- New bounds linked to high-dimensional sphere packing
- Better limits for binary and spherical error-correcting codes
- A construction showing the existence of non-sofic groups
- A claimed disproof of Connes’s rigidity conjecture
- New lower bounds in arithmetic circuit complexity
- A quantum parallel repetition theorem for general two-player games
- New hardness results for the closest vector problem
- A result on Ehrhart’s volume conjecture
- New bounds for multicolour Ramsey numbers
- Results covering two extremal graph theory conjectures
One of the most striking claims concerns non-sofic groups. Mathematicians have long asked whether every countable group can be approximated through finite permutations. The Astra-generated paper presents an explicit construction that says the answer is no.
Why the findings need independent checking
A Lean certificate can catch gaps in formal logic once a proof has been translated correctly. Researchers still need to examine the assumptions, definitions, novelty and wider meaning of every result.
OpenAI has taken responsibility for the claimed correctness of the manuscripts and says the core arguments came from its AI system. Peer review and attempts by independent mathematicians to reproduce the work will now determine whether Astra has delivered ten lasting research advances.