Humanity's Last Exam is a benchmark of approximately 3,000 expert- level questions spanning mathematics, physics, biology, law, philosophy, and other domains. The questions were contributed by domain experts and filtered for difficulty — a question only entered the benchmark if GPT-4 and several other frontier models failed to answer it correctly.
HLE was designed as a successor to MMLU that would not saturate quickly. When introduced, frontier models scored in the low single digits of percent. As of 2026, top models including Claude Opus 4.6 and Gemini 3.1 Pro have crossed 50 percent on HLE, which was not expected until at least 2027.
The rapid progress on HLE has recalibrated expectations about the frontier of AI capability. It has also renewed debate about what benchmarks actually measure. HLE is a knowledge benchmark, not a reasoning benchmark, and high scores on it do not automatically translate to capability at novel problems outside the expert- question format.