From Olympiad Problems to Navier-Stokes: How Quickly Is AI Getting Better at Mathematics?

Very quickly: between July 2024 and July 2025, AI systems went from silver-medal-standard performance on the International Mathematical Olympiad to gold-medal-standard performance. Then, on September 8, 2026, OpenAI published an AI-generated proposed resolution of the Navier–Stokes Millennium Prize Problem with a Lean formalization—an enormous claim, but one that is far too fresh to treat as settled mathematics.

The useful engineer’s answer is that this is not a smooth “LLMs got 20% smarter” curve. It’s a rapid shift from generating plausible solutions, to searching many possible proof paths, to checking results mechanically, to running coordinated agent systems with research-scale budgets. That stack is what’s starting to matter.

The timeline is real, but the tasks changed

In July 2024, Google DeepMind reported that AlphaProof and AlphaGeometry 2 scored 28 out of 42 on that year’s IMO: four of six problems, at the top of the silver-medal range. This was already notable, but it was a specialized pipeline. Experts translated the problems into formal languages, and some successful runs took up to three days. AlphaProof searched for Lean proofs; AlphaGeometry handled a geometry-heavy slice of the workload.

A year later, an advanced Gemini Deep Think system received a 35/42 score on the 2025 IMO, reaching gold-medal standard. The important operational change was not just the extra seven points. It worked end-to-end from natural-language problem statements, produced natural-language proofs, and did so within the competition’s 4.5-hour time limit. That’s a meaningful jump in usability: fewer handcrafted translations and less dependence on a domain-specific frontend.

Still, don’t flatten these results into “the model is now a mathematician.” Olympiad problems are exceptionally hard, but they are short, carefully specified, and have a known answer space. Research mathematics is messier. Definitions drift across papers, useful lemmas are missing, notation is inconsistent, and deciding which subproblem matters can be harder than executing the proof.

When you see a math-capability graph, ask what changed besides the base model: the benchmark, available tools, wall-clock limit, number of parallel attempts, human hints, formalization effort, and grading procedure. Those details are the difference between an interesting demo and a durable capability.

Why Navier–Stokes is a different class of evidence

The three-dimensional Navier–Stokes existence-and-smoothness problem asks, roughly, whether an initially smooth incompressible fluid can develop a finite-time singularity despite viscosity. It is a Millennium Prize Problem because the answer has resisted mathematical attack for decades. A valid counterexample with the right properties is not merely a difficult exercise; it changes the known landscape of a major field.

OpenAI’s September 8 release says an internal multi-agent system found such a counterexample: an initially stationary fluid, acted on by a smooth external force, develops a finite-time singularity while maintaining finite energy. The company says the construction establishes statements C and D in Clay’s official formulation. It also released a conventional writeup and a Lean formalization.

That last part matters a lot. Lean is not a vibe check for prose. A proof assistant reduces a theorem to small logical steps that a compact trusted kernel can verify. If the formal statement correctly captures the intended Navier–Stokes claim, and the Lean development checks against the intended dependencies, it gives a much stronger correctness guarantee than a fluent model-generated PDF.

But “if” is doing real work. Formal verification proves the theorem you encoded, not automatically the theorem you meant to encode. Mathematicians still need to inspect the assumptions, regularity conditions, function spaces, forcing term, boundary behavior, and translation from the Clay formulation into Lean. They also need to judge whether the explanatory proof exposes a genuine mechanism rather than an artifact hidden by an overly narrow specification.

The biggest capability jump is the harness

OpenAI’s disclosed setup is a strong reminder that frontier mathematical performance is now a systems problem. The company says its successful Navier–Stokes effort involved on the order of 10,000 concurrent agents, 2.7 million messages, roughly 130 billion output tokens, code execution, access to a cached internet, cross-pollination between groups, and 17 additional hours for Lean formalization and verification. The claimed proof arrived about 88 hours after the work began.

That does not mean a single chat window can solve any famous open problem in a weekend. It means a very capable underlying model, given massive parallel search and a tight feedback loop, may produce research outcomes much faster than an individual can. The unit of progress is increasingly the model-plus-orchestrator-plus-tools system.

There’s a familiar software-engineering analogy: a good code generator is useful, but CI, tests, static analysis, reproducible builds, code review, and issue triage turn it into something a team can trust. Formal mathematics is building an unusually strict version of that pipeline. The generator proposes lemmas and proof strategies; search explores branches; numerical tools test intuitions; a proof assistant makes correctness cheap to re-check.

How to evaluate the next “AI solved math” headline

Use a stricter review standard than you would for a benchmark announcement. You don’t need a PDE PhD to separate compelling evidence from marketing.

1. Check the exact claim. Is it a solution to the original open problem, a result for a simplified variant, a numerical observation, or a new family of examples in a related equation?
2. Look for independent review. A public paper and source code are necessary; scrutiny by domain experts who were not part of the announcement is what turns a claim into accepted mathematics.
3. Check whether a formal proof exists, what theorem it verifies, and whether the formalization is publicly reproducible. A Lean repository is stronger than a screenshot saying “verified.”
4. Inspect the compute and harness. One model pass, a thousand-agent search, and a human-guided formalization are all useful results—but they measure different capabilities.
5. Separate discovery from explanation. Can researchers understand the new technique, reuse it, generalize it, or derive new conjectures from it?

What this changes for engineers now

Don’t wait for autonomous theorem discovery to become routine before changing how you work. The near-term product is a research copilot that can turn a vague technical question into candidate derivations, symbolic experiments, simulations, citations to inspect, and—where the domain permits it—machine-checkable artifacts.

Start with problems that have executable feedback. Ask a model to derive an invariant, then verify it with property tests. Have it propose an algorithmic optimization, then benchmark it. Let it write a proof sketch for a correctness property, then encode the critical claim in Lean, Coq, or a model checker if the stakes justify it. Treat the model’s answer as a search proposal and your verification pipeline as the authority.

The headline pace is extraordinary: silver-standard Olympiad performance in 2024, gold-standard performance in 2025, and a publicly released but still-unsettled Navier–Stokes resolution claim in September 2026. The more durable lesson is less sensational: AI is getting much better at mathematics when we give it the same thing we learned to give software—tight specifications, lots of parallel work, fast feedback, and a way to reject results that merely look correct.