What it measures
Algorithmic problem-solving on problems posted after a model's training cutoff (contamination-free).
How it works
Problems are scraped from public coding judges. Each problem records its publication date so evaluations can be filtered to problems newer than the model's knowledge cutoff. Models output a solution; the original judges' tests score it.
Caveats
- Competitive programming is a narrow slice of coding. High LiveCodeBench score does not imply strong engineering ability.
- Time pressure and unusual algorithmic tricks dominate the problem distribution.
- Score depends heavily on which date window is selected.
Primary source
Related benchmarks
- HumanEval — 164 Python coding problems with hidden unit tests.
- SWE-Bench Verified — 500-instance cleaned subset of SWE-Bench with verified test reliability.
How we read this benchmark
Our Benchmark Methodology explains which benchmarks we cite on model pages, the caveats we apply, and why we never present any single benchmark as a verdict.