OpenAI's AI Cracks Ten Open Maths and Computing Problems

The News

OpenAI has published a research note titled "Ten advances in mathematics and theoretical computer science," claiming that an internal version of its next major model, referred to as Astra, contributed to fresh progress on ten long-standing open problems. The company framed the work around a simple ambition: "We want to empower scientists and mathematicians with tools that accelerate discovery."

The ten results span several hard corners of pure mathematics and theory. They include high-dimensional sphere packing, binary and spherical codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, the closest vector problem, Ehrhart's volume conjecture, multicolor Ramsey numbers and a set of extremal number conjectures.

Alongside the research, OpenAI said it is opening up ChatGPT access to 100,000 academic researchers. The company put the raw compute cost of finding one of the solutions at roughly "$2,000 at Sol API rates," a figure that hints at how cheap machine-assisted discovery is becoming for problems that have resisted human effort for years.

Why It Matters

The announcement matters less for any single theorem than for what it signals: frontier models are moving from summarising known mathematics to generating arguments that specialists had not found. That is a different order of claim. When GPT-4 launched in March 2023, the debate was whether a model could reliably do school-level arithmetic. Barely three years on, the conversation has shifted to whether a system can meaningfully attack conjectures at the research frontier.

OpenAI was careful on the question of credit. The note argued that presenting a proof produced entirely by a machine as human work would misrepresent both the system's contribution and the character of genuine intellectual labour. That framing tries to head off the obvious backlash from a research community that prizes authorship above almost everything.

There is reason for caution too. Extraordinary mathematical claims demand independent checking, and history is full of announced proofs that unravelled under peer review. The real test is not the press note but whether working mathematicians can verify, formalise and build on the arguments over the coming months.

Indian Angle

India has an outsized stake in exactly the fields on this list. Combinatorics and theoretical computer science, the home of Ramsey numbers and extremal conjectures, have been strengths at the Indian Statistical Institute, the Tata Institute of Fundamental Research, Chennai Mathematical Institute and the IITs for decades. If machine-assisted proofs become routine, Indian departments that are rich in talent but thin on funding could gain a genuinely new instrument.

The economics are the sharpest angle. A compute bill of around $2,000, roughly ₹1.7 lakh, to make progress on a problem that stumped experts is trivial for a well-funded lab and no longer absurd for an Indian university grant. That flips the usual constraint: the scarce input becomes the research question, not the money. It also raises a policy question for MeitY and the IndiaAI mission, which have poured public money into compute and datasets but said little about how domestic researchers should use foreign frontier models for original science.

There is a competitive lens as well. India's home-grown model builders, Sarvam and Krutrim among them, are focused on language, voice and Indian-context applications, not on frontier reasoning. This announcement widens the gap between applied Indian AI and the deep-reasoning frontier held by a handful of American labs, and sharpens the question of where Indian institutions should place their bets.

FAQ

What exactly did OpenAI claim to solve?

OpenAI said an internal model contributed to progress on ten open problems across mathematics and theoretical computer science, including sphere packing, multicolor Ramsey numbers, Connes's rigidity conjecture and the closest vector problem. It describes these as advances, not fully settled and peer-reviewed results.

Which model was used?

The company attributed the work to an internal version of a next-generation model it refers to as Astra, rather than the standard public ChatGPT release available to consumers.

How much did it cost?

OpenAI cited a compute cost of about "$2,000 at Sol API rates" for reaching one of the solutions, roughly ₹1.7 lakh, illustrating how affordable machine-assisted mathematical exploration is becoming.

What should Indian researchers do with this?

Watch for independent verification before treating any result as settled, and consider how affordable model access could support original combinatorics and theory work at institutions such as ISI, TIFR and the IITs.

This story was reported by OpenAI. Read the full original coverage at OpenAI.