ConjectureBench: 15k open math problems for you to solve

People and frontier models are starting to solve math problems that have been open for decades. Last week, a mathematician working with Claude Fable 5 found a counterexample to the Jacobian conjecture, which had been open since 1939 (link). GPT-5.6 also produced a proof of the cycle double cover conjecture, open since the 1970s, which has since been independently verified (link).

We built ConjectureBench to help more people do this as a fun two day sprint experiment. It brings 14,865 open math problems together in one place, with a link back to the source for each one. Think of it as a treasure chest where anyone, human or model, can pick a problem and start.

If you find a solution or improve on a known result, please share it with the community! Open a claimed-result GitHub issue. Include the problem ID, the people and models involved, a link to the result, and any validation. We will add it to the catalog so others can find and evaluate it. You can also submit source-backed corrections by pull request.

We hope ConjectureBench gives people and models more problems to explore. We cannot wait to see what you find.

See more at https://github.com/bespokelabsai/conjecture-bench