A general-purpose AI model has reportedly solved a problem that stumped mathematicians for four decades. Not a narrow, purpose-built system trained exclusively on proofs. A general-purpose model, the kind people use to draft emails and argue about pizza toppings.

Mathematician Ernest Ryu used OpenAI’s GPT-5 to tackle a 40-year-old open problem in convex optimization, producing a solution that is now undergoing formal verification. If the proof holds up, it would mark one of the most significant demonstrations yet of AI’s capacity to contribute genuine mathematical discoveries, not just assist with them.

From competition tricks to research-grade breakthroughs

The line between “AI can do math homework” and “AI can do math research” has been blurring for a while now. But recent developments suggest we may have crossed it entirely.

OpenAI’s internal reasoning models have produced over 10 new solutions to Erdős-style combinatorics problems. For the uninitiated, Paul Erdős was one of the most prolific mathematicians in history, and the problems bearing his name are notoriously difficult open questions in combinatorial mathematics. Some of these AI-generated solutions are reportedly being considered for publication in top journals.

In English: the AI isn’t just checking answers. It’s generating novel mathematical results that peer reviewers at elite publications consider worth publishing.

This follows a broader trend of AI systems performing at elite human levels in structured mathematical competitions. At the 2025 International Mathematical Olympiad, models from both OpenAI and Google DeepMind achieved gold-medal performance by solving 5 of 6 problems under standard competition constraints. These aren’t trivial exercises. The IMO is widely considered the most prestigious math competition in the world, and gold medals typically go to the top fraction of competitors from dozens of countries.

Google DeepMind’s AlphaEvolve system has been making its own waves. When tested against 67 research-level mathematical problems, it improved upon the best known results in 23 of them. That’s a hit rate of roughly one in three on problems that professional mathematicians have been working on for years.

Why convex optimization matters beyond pure math

Convex optimization isn’t the kind of math that trends on social media, but it quietly underpins enormous swaths of modern technology. Machine learning training algorithms, financial portfolio construction, signal processing, and logistics planning all rely on it. A 40-year-old open problem in this field isn’t an obscure curiosity. It’s a bottleneck that has shaped the boundaries of what’s computationally practical.

Ryu’s work with GPT-5 is particularly notable because the model used was general-purpose. Previous AI breakthroughs in mathematics, like DeepMind’s work on knot theory or its AlphaGeometry system, involved models specifically architected or fine-tuned for mathematical reasoning. GPT-5, by contrast, is designed to handle everything from creative writing to code generation to customer service chatbots.

The implication is uncomfortable for anyone who assumed that genuine mathematical creativity would remain a uniquely human domain for the foreseeable future. If a model built to be a jack-of-all-trades can crack problems that specialists couldn’t solve in 40 years, the ceiling for what purpose-built mathematical AI systems might achieve becomes very hard to estimate.

Mathematicians themselves are beginning to reframe their relationship with these tools. There’s growing interest in what researchers call “autoformalization,” the process of translating informal mathematical reasoning into formal, machine-verifiable proofs. The vision is not AI replacing mathematicians but AI formalizing significant portions of mathematics, turning intuitions and sketches into rigorous, verified results at a pace humans simply can’t match.

What this means for crypto and beyond

Here’s the thing. Cryptography is applied mathematics. Every encryption scheme, every zero-knowledge proof system, every consensus mechanism rests on mathematical assumptions about what problems are hard to solve. When AI gets dramatically better at solving hard math problems, the entire security landscape shifts.

This doesn’t mean your Bitcoin wallet is getting cracked tomorrow. The problems AI is solving right now are in optimization and combinatorics, not in the number-theoretic domains that underpin most cryptographic systems. But the trajectory matters. If general-purpose models are solving 40-year-old open problems today, the question of when they’ll be applied to cryptographic assumptions becomes a matter of when, not if.

On the constructive side, enhanced AI mathematical reasoning could accelerate the development of new cryptographic protocols. Zero-knowledge proof systems, homomorphic encryption schemes, and post-quantum cryptographic standards all require deep mathematical innovation. AI systems that can generate novel proofs and improve upon existing results could compress years of research into months.

The formal verification angle is equally relevant to crypto. Smart contract auditing, protocol correctness proofs, and security guarantees all benefit from the kind of rigorous formalization that AI is increasingly capable of producing. A world where AI can translate informal mathematical reasoning into verified proofs is a world where protocol security becomes more provable and less dependent on “we looked at it really hard and didn’t find any bugs.”

For investors watching the AI-crypto intersection, the signal is clear. Companies and protocols building infrastructure for formal verification, AI-assisted auditing, and post-quantum cryptography are positioning themselves for a future that’s arriving faster than most roadmaps anticipated. The mathematical capabilities demonstrated by GPT-5 and AlphaEvolve aren’t just academic milestones. They’re early indicators of a fundamental shift in what’s computationally possible, and crypto sits directly in the blast radius of that shift.