ComBench is a new benchmark designed to evaluate the combinatorial reasoning abilities of large language models, particularly in Olympiad-level mathematics. Current models show significant gaps in creative mathematical reasoning, with the strongest achieving only moderate performance on the benchmark's problems.