LiveBench

What It Measures

LiveBench is a contamination-resistant benchmark covering a broad range of capability areas: math, reasoning, coding, instruction following, language, and data analysis. The defining feature is its monthly refresh cadence — questions are added monthly from current sources (recent news, recent academic papers, recently-modified open-source codebases), and older questions are rotated out. This makes LiveBench substantially harder to game through training-data inclusion than fixed benchmarks, since the test set is constantly moving forward in time.

As traditional benchmarks like MMLU and HumanEval became saturated and contamination-prone in 2024-2025, LiveBench emerged as one of the more credible alternatives for evaluating frontier model capability. The benchmark covers enough capability areas that a strong LiveBench score is a meaningful general-purpose intelligence signal, while the monthly refresh makes the leaderboard particularly informative as models are released.

How It Works

Each month, new questions are sourced from the previous month's events and content. The test set rotates: questions older than a fixed window are removed, and the new questions are added. This means a 'LiveBench score' has an implicit timestamp — the score reported is on the test set that was current at that month's evaluation. Scores from different time periods are not directly comparable, though the benchmark publishes month-over-month trend data.

The scoring methodology aggregates results across the capability areas (math, reasoning, coding, etc.) into both a composite score and per-category scores. For comparing models, the composite score is most commonly cited, though per-category scores can reveal model-specific strengths and weaknesses.

Current Leaders