ASTRA: OpenAI's Next Model Just Solved 10 Open Mathematical Problems

OpenAI has just announced ASTRA, its next major model, in an unusual way: tucked into an apparently academic blog post titled "Ten advances in mathematics and theoretical computer science". No press conference, no public demo, no release date—just a result that speaks louder than any keynote: ten open problems, some stagnant for half a century, have just been solved by an internal version of ASTRA.

Whether the model will eventually be released as GPT-6 or as a high-end variant of the GPT-5 line, OpenAI has not yet disclosed. What it does assert, however, is that the era of the model-tool is giving way to the era of the model-researcher.

Ten problems, several decades of stagnation

The ten results are not "difficult exercises." They are conjectures and open problems that the mathematical community has been grappling with—mostly for more than ten years, often much longer. The affected fields include:

• high-dimensional geometry
• coding theory
• arithmetic circuit complexity
• group theory
• operator algebras
• quantum complexity
• lattice-based cryptography
• extremal combinatorics

Among the most striking results announced by OpenAI:

• the first explicit non-sofic group ever constructed;
• the refutation of Connes' rigidity conjecture;
• quantum parallel repetition demonstrated for general entangled two-player games;
• the Ehrhart volume conjecture proved;
• the first improved exponent for sphere packing in general since 1978.

A half-century without progress on sphere packing. A model that, in an afternoon, pushes the boundary. That's the announcement.

Not just ideas: formalized proofs

The part that should make anyone still doubting the scientific value of LLMs think: ASTRA did not produce "heuristics". The model generated the central arguments, then formalized them in Lean—a formal proof language that produces machine-verifiable certificates.

In practice, each theorem is accompanied by a Lean file that a compiler can verify bit by bit. There is no "you can see here that..." to interpret—the proof is either correct or it is not. It is. A 249-page manuscript accompanies it, written and structured like a research paper.

This is a change in nature. A model that writes plausible text on a mathematical subject has existed since GPT-4. A model that writes a mechanically verified proof of an open problem since 1978—that's new.

The cost: less than $2,000

OpenAI provides an almost provocative figure: less than $2,000 combined to generate the ten proofs, priced at the "Sol" API rate (the internal cost reference for the GPT-5.6 line). In other words: where a PhD student costs €30,000 per year and a senior researcher much more, the entry barrier to attempt solving a problem deemed impossible drops to a few dozen euros per try.

The economic landscape of fundamental research is transformed. And not just in math: as soon as a field has a formal proof or verification system (theoretical physics, formal program security, cryptography, structural bioinformatics), the same pattern can replicate.

What ASTRA concretely changes

1. Fundamental research becomes tractable for non-mathematicians

An engineer with a $500 budget and a good question can, in principle, attempt to tackle a problem long reserved for specialists. What matters becomes the formulation rather than raw technical ability.

2. Peer review will need to reinvent itself

A 249-page manuscript produced in a few hours, with Lean proofs provided—how does a human review committee validate that, in what time frame, with what tools? The mathematical community will need to industrialize machine verification. Journals that accept a Lean certificate as an equivalent of a written proof will gain years of advantage.

3. The line between "model" and "researcher" blurs

ASTRA is no longer an assistant. It is an autonomous research agent that poses conjectures, proves them, and publishes them in verifiable form. The ethical question—who "signs" a paper whose brain is a model—becomes urgent.

When will ASTRA be released?

OpenAI is not committing. The model could be commercialized as GPT-6, or as a premium evolution within the GPT-5 line (perhaps GPT-5.7 ASTRA, logically following their nomenclature). No date is communicated.

But the announcement, in its form, sends a perfectly clear signal to the competition: Google, Anthropic, Moonshot AI, DeepSeek—the bar has just been raised. The next battleground for frontier models is no longer the MMLU benchmark or the general public multimodal agent. It is original, verifiable, low-cost scientific production.

And OpenAI, with ASTRA, has just fired the first shot.