🧮 AI solves math problem that researchers failed to crack for six years

AI Solves Unsolved Math Problem, Signaling a New Era in Scientific Discovery

For the first time, an artificial intelligence system has successfully solved a problem from FrontierMath: Open Problems, a challenging benchmark of real-world mathematical research questions that have stumped human mathematicians. This breakthrough, achieved by GPT-5.4 Pro, marks a significant milestone in the evolving relationship between AI and scientific exploration.

A Problem Years in the Making

The solved problem originated as a conjecture from mathematician Will Brian in a 2019 paper co-authored with Paul Larson. Despite numerous attempts by Brian, Larson, and other experts, the problem remained unresolved. Brian categorized the challenge as “Moderately Interesting” within the FrontierMath framework, highlighting its complexity and potential impact.

From Conjecture to Potential Publication

Brian now intends to prepare the solution for publication in a specialist journal. He anticipates that the AI-generated solution may similarly inspire new research avenues, potentially incorporating follow-on work stemming from the model’s insights. Kevin Barreto and Liam Price were the first to elicit the solution from GPT-5.4 Pro and have been offered co-authorship on the resulting paper, alongside Brian. Geby Jaff also independently elicited a solution shortly after.

Beyond GPT-5.4 Pro: A Chorus of AI Success

The initial success with GPT-5.4 Pro wasn’t an isolated incident. Epoch AI, the organization behind FrontierMath, replicated the solution using its own testing framework. Remarkably, other advanced AI models – Gemini 3.1 Pro and Claude Opus 4.6 (max) – also demonstrated the ability to solve the problem, albeit with varying degrees of consistency.

The Rise of AI in Mathematical Research

This achievement builds upon a growing trend of AI capabilities in mathematics. Data from Epoch AI shows a dramatic increase in performance on FrontierMath problems. GPT-4 achieved a roughly 5% success rate in 2024, while GPT-5.4 Pro jumped to 50% in March 2026. This exponential growth suggests that AI could become an increasingly valuable tool for mathematicians.

From Competition Math to Research Assistant

The evolution of AI’s mathematical prowess is also shifting its focus. In 2025, AI began tackling competition-level math problems, comparable to those found in the International Mathematical Olympiad (IMO). By 2026, AI is demonstrating the ability to contribute to research-assistant-level mathematics across various applied domains.

Implications for the Future of Scientific Discovery

The successful resolution of this FrontierMath problem has broader implications for scientific discovery. It suggests that AI can not only automate routine tasks but also assist in tackling complex, open-ended research questions. This could accelerate progress in fields beyond mathematics, including physics, chemistry, and biology.

Did you realize? A full transcript of the conversation with GPT-5.4 Pro leading to the solution is publicly available on the FrontierMath website, offering a unique glimpse into the AI’s reasoning process.

The Role of “Elicitation”

The process of obtaining the solution wasn’t simply a matter of asking the AI the question directly. It involved “elicitation” – carefully crafting prompts and guiding the AI’s thinking to arrive at the correct answer. This highlights the importance of human-AI collaboration in pushing the boundaries of knowledge.

FAQ

Q: What is FrontierMath: Open Problems?
A: It’s a benchmark of real, unsolved mathematical research problems created by mathematicians.

Q: Which AI model first solved the problem?
A: GPT-5.4 Pro was the first to elicit a confirmed solution.

Q: Will the solution be published?
A: Yes, Will Brian plans to write up the solution for publication in a specialist journal.

Q: What does this mean for the future of math research?
A: It suggests AI can be a powerful tool for assisting mathematicians and accelerating scientific discovery.

Pro Tip: Explore the FrontierMath website to learn more about the benchmark and the solved problem. You can even view the original chat transcript with GPT-5.4 Pro!

Seek to stay informed about the latest advancements in AI and scientific discovery? Subscribe to our newsletter for regular updates, and insights.