Something unusual is happening in the world of pure mathematics, and it isn’t coming from a university lab. OpenAI says an internal AI system has helped crack ten mathematical and computer-science problems that have sat unsolved for at least a decade, in some cases far longer. The announcement, paired with a separate initiative opening advanced ChatGPT access to 100,000 researchers, signals that AI advances in mathematics are moving from novelty to genuine research infrastructure.
Summary
Key takeaways
- OpenAI launched ChatGPT for Academic Researchers, giving 100,000 scientists and mathematicians free access to its top ChatGPT models.
- An unreleased OpenAI model produced an AI-generated disproof of the Erdős unit-distance conjecture in May.
- Ten new results, achieved with an internal version of a next-generation model called Astra, solve open problems across eight fields of mathematics and theoretical computer science.
- Finding the solutions cost roughly $2,000 in tokens at Sol API rates, according to OpenAI.
- Human researchers turned the AI-generated arguments into manuscripts, and the model then formalized each proof in a Lean certificate.
OpenAI opens ChatGPT access to 100,000 academic researchers
OpenAI’s starting point for this push is access, not just output. The company recently rolled out ChatGPT for Academic Researchers, an initiative that hands 100,000 scientists and mathematicians free use of its most capable ChatGPT models. The idea, as OpenAI frames it, is to put stronger reasoning tools directly into the hands of people working on unsolved problems, rather than keeping the most advanced models behind a paywall reserved for enterprise customers.
That matters because access has often been the bottleneck for AI-assisted research. Academic budgets rarely stretch to cover premium AI subscriptions at scale, and individual researchers testing frontier models on niche mathematical questions is not something most institutions can fund broadly. By removing that barrier for a large cohort of scientists, OpenAI is effectively betting that wider access will surface more of the kind of results it’s now showcasing.
A quiet breakthrough: AI disproves the Erdős unit-distance conjecture
Back in May, OpenAI shared something that hinted at what was coming: an AI-generated disproof of the Erdős unit-distance conjecture, a problem tied to the mathematician Paul Erdős that had stumped researchers for years. The disproof surfaced almost by accident, discovered while OpenAI was evaluating an unreleased model rather than during a dedicated research push.
That single result turned out to be more than a one-off curiosity. According to OpenAI, the Erdős disproof has already inspired further developments in mathematics and theoretical computer science, with subsequent papers building on the approach. This is one of the clearer “why it matters” moments in the story: when an AI-generated result triggers follow-on work from human mathematicians, it suggests these systems aren’t just producing isolated answers — they’re contributing techniques other researchers can extend.
Ten new AI-driven solutions to problems mathematicians couldn’t crack for decades
The centerpiece of OpenAI’s announcement is a batch of ten new results, all tackling problems that had seen no progress on their main question for at least ten years, and in many cases considerably longer. OpenAI describes these problems as being of substantial interest to their respective mathematical communities, with several considered important across mathematics as a whole.
Scope and diversity of solved problems
The ten results span a wide range of technical territory: high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. That breadth is itself notable — these aren’t adjacent sub-questions in a single niche, but distinct open problems from separate corners of mathematics and computer science, each with its own decades-long research history.
Among the specific advances OpenAI listed:
- New upper bounds on sphere-packing density in high dimensions, reaching down to the Cohn–Elkies threshold.
- Exponentially improved bounds on the maximum size of binary and high-dimensional spherical codes at a given minimum distance.
- A construction proving the existence of non-sofic groups, resolving a central open question in group theory.
- A disproof of Connes’s rigidity conjecture, which held that certain groups are uniquely determined by their von Neumann algebras.
- New lower bounds for computing the permanent using arithmetic circuits, including an arithmetic-formula lower bound on the order of n^4/log n.
- An exponential parallel repetition theorem for general two-player quantum games, extending a classical complexity principle into the quantum setting.
- Polynomial-factor hardness of approximation for the closest vector problem, a lattice question relevant to post-quantum cryptography.
- A resolution, in every dimension, of the maximum volume of a convex body whose centroid is its only interior lattice point — part of Ehrhart’s volume conjecture.
- A superexponential lower bound for multicolor triangle Ramsey numbers, resolving Erdős problem 183.
- Results on compactness and degeneracy conjectures in extremal graph theory, resolving Erdős problems 146 and 180.
Technical process and cost efficiency using the Astra model
All ten results were produced by an internal version of Astra, described by OpenAI as its next major model. What stands out is the cost: OpenAI estimates that the total token usage needed to find solutions to all ten problems would run roughly $2,000 at Sol API rates. For research that took human mathematicians a decade or more to make no progress on, a few thousand dollars in compute is a strikingly small price tag — and it’s one of the clearest signals of how AI advances in mathematics are shifting the economics of tackling hard open problems.
The process wasn’t fully automated, though. Human researchers took the model’s raw arguments and turned them into structured manuscripts, working alongside the same model. After that, the model itself formalized each argument into a Lean certificate — a machine-checkable proof format used to verify mathematical claims with rigor. OpenAI also released a narration of the model’s thinking process for each solution, giving outside researchers a window into how the system reasoned its way to each answer.
That combination — AI-generated arguments, human-prepared manuscripts, and machine-formalized Lean certificates — points to a workflow where AI systems generate the raw mathematical insight while formal verification tools and human oversight handle confirmation. It’s a division of labor that could become a template as more of these AI-assisted results start appearing across other open problems.
Why this matters for research and competition
The broader implication here goes beyond ten solved problems. If a single internal model can produce publishable-grade results across eight distinct mathematical disciplines for roughly $2,000 in compute, the constraint on tackling long-standing open problems starts shifting from mathematical insight toward access to capable models — which is exactly what OpenAI’s academic access program is designed to expand. That combination of cheap computation and wide researcher access is likely to draw close attention from rival AI labs and from mathematics departments weighing how to integrate these tools into their own work.
OpenAI also notes that this line of work has already spurred additional research building on the Erdős conjecture disproof, with several follow-up papers exploring related questions in sum-product theory, incidence geometry, and computational complexity. Whether that momentum continues at the same pace, or whether other AI labs produce comparable results with different models, will likely shape how quickly AI-assisted proof generation gets absorbed into everyday mathematical practice.
FAQ
What is ChatGPT for Academic Researchers?
It’s an initiative by OpenAI providing 100,000 scientists and mathematicians free access to advanced ChatGPT models to accelerate discovery.
Which longstanding mathematical problem did OpenAI’s AI model disprove?
An AI-generated disproof of the Erdős unit-distance conjecture was produced using an unreleased OpenAI model, first shared in May.
What types of problems did Astra solve with AI assistance?
Astra, OpenAI’s internal next-generation model, produced results in high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics.
How were the AI-generated proofs validated?
Human researchers prepared manuscripts from the AI-generated arguments, and the model then formalized each proof in a Lean certificate for machine-checkable verification.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
