ModelRefs / LiveCodeBench Leaderboard — AI Model Scores
LiveCodeBench Leaderboard — AI Model Scores
Fresh competitive programming problems with timestamped contamination guard. Current leaders, methodology, and citation sources for LiveCodeBench.
Overview
Fresh competitive programming problems with timestamped contamination guard.
How it is measured: pass@1 on problems released after model training cutoff.
What this benchmark measures
- competitive-programming code generation
- code execution reasoning
- test-output prediction
- bounded code self-repair
Relevant to:
- coding-model screening
- code-generation regression testing
- time-windowed coding evaluation portfolios
Failure modes it exercises:
- incorrect executable solutions
- failure on recently published contest problems
- code-execution and output-prediction errors
Method and its limits
Continuously collect problems from LeetCode, AtCoder, and Codeforces and evaluate code generation plus additional code-related scenarios under a pinned release, task window, language, harness, and sampling protocol.
- Scores are sensitive to benchmark release, problem date window, scenario, language, model cutoff claim, sampling, and harness revision.
- A contamination-mitigation design does not prove that a particular model has not seen a problem or related solution.
- Competitive-programming correctness does not establish repository-scale engineering, secure coding, maintainability, or agent-tool reliability.
Dataset
- Dataset
- LiveCodeBench
- Type
- continuously collected competitive-programming problems and code-related evaluation scenarios
- Freshness
- current
The official repository documents continuously collected contest problems and multiple coding scenarios. Every result must identify the benchmark release or task window because the corpus evolves over time.
How to read this score
Compare results only when scenario, release or task window, language, pass@k or sampling setup, and harness match. Treat the score as system performance on that slice, not a general software-engineering rank.
Similarity to real tasks: Medium — The tasks require executable solutions to real contest problems, but they do not reproduce repository context, product requirements, code review, deployment, or maintenance work.
Data contamination risk: Low — The benchmark's time-windowed contest-problem design reduces exposure when a result pins both model cutoff and task window. This is not a blanket contamination-free guarantee for every model row.
Benchmark gaming risk: Medium — Public tasks, repeated leaderboard use, benchmark-specific prompting, sampling, scenario selection, and harness tuning can affect results.
What you still need to test yourself
- Test representative repositories, languages, dependencies, security constraints, review quality, and tool or agent scaffolds.
- Record task window, model snapshot, prompt, sampling, execution environment, latency, cost, and repeated-run variance.
This benchmark supports decisions about:
- Screen coding candidates on a pinned, time-bounded competitive-programming protocol.
- Identify code-generation and execution-reasoning failures that should enter an internal coding regression suite.
Limitations
- LiveCodeBench does not establish repository-scale engineering quality, security, maintainability, or production readiness.
- Source-scoped rows from different releases, windows, scenarios, or harnesses must not be normalized into one directly comparable score.
- Model-specific contamination, prompt, sampling, language, and harness parity remain run-level evidence requirements.
Sources re-reviewed 2026-07-11. LiveCodeBench is treated as current only when a score pins the benchmark release or task window; the 2024-06-06 paper revision anchors this methodology review, not every future leaderboard row.
Sources
- LiveCodeBench official repository LiveCodeBench authors · accessed 2026-07-11
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code LiveCodeBench authors · accessed 2026-07-11
How this benchmark is scored
| Category | coding |
|---|---|
| Maximum score | 100 pass@1 |
| Direction | Higher is better |
Primary source: https://livecodebench.github.io/
Published results
| Model | Score | Run date | Source |
|---|---|---|---|
| GPT-5 | 72 | 2026-05-01 | Aggregated public reports |
| DeepSeek R1 | 65.9 | 2026-05-01 | Aggregated public reports |
Each score reflects the protocol and date of its own source run. Results from different harnesses are not directly comparable.
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to LiveCodeBench Leaderboard — AI Model Scores.