Erdős Problems Emerge as New Frontier Math Benchmark for AI Models

Researchers are turning to unsolved Erdős problems as AI math benchmarks — and OpenAI reportedly used an unreleased internal model, not GPT-5.5, to tackle them.