OpenProblemBench: when presented with 82 open problems in mathematics and theoretical physics, GPT-6 Astra solves 14 per cent of them

A new benchmark tests AI models against 82 unsolved problems from the literature. GPT-6 Astra achieves a 14 per cent success rate for solutions deemed valid, more than double that of the best open-source models.

The usual test benches check what the models already know. On the contrary, OpenProblemBench aims at what no one knows yet: 82 unresolved problems from literature in mathematics and theoretical physics. Each is provided with its research context, assumptions and advances already achieved. The authors have selected issues whose proposed solutions lend themselves to a relatively clear verification of their decisive mathematical or numerical statements.

Judges without correction

In the absence of a reference solution, four evaluation models independently note the accuracy, completeness and progress of each response. Of the seven configurations tested, GPT-6 Astra achieved the highest rated average resolution rate, 14%, compared to 5.5-6.7% for the large open models evaluated and 2.4-3.7% for the flash versions.

The case analysis shows that the best answers reformulate the problem, put forward general arguments that are not limited to numerical checks, and demonstrate the steps needed to conclude. The authors present the benchmark as a ground for measuring what AI can really bring to fundamental theoretical science, and where it still ends.