LiveCodeBench v6 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | qwen3-7-max | 0.916 |
| 2 | namazu | 0.9033 |
| 3 | kimi-k2-6 | 0.896 |
| 4 | qwen3-7-plus | 0.896 |
| 5 | nemotron-3-ultra-550b-a55b | 0.89 |
| 6 | seed-2-0-pro | 0.878 |
| 7 | qwen3-6-plus | 0.871 |
| 8 | step-3-5-flash | 0.864 |
| 9 | kimi-k2-5 | 0.85 |
| 10 | glm-4-7 | 0.849 |
| 11 | qwen3-6-27b | 0.839 |
| 12 | qwen3-5-397b-a17b | 0.836 |
| 13 | kimi-k2-thinking-0905 | 0.831 |
| 14 | gpt-oss-120b-high | 0.819 |
| 15 | seed-2-0-lite | 0.817 |
| 16 | k-exaone-236b-a23b | 0.807 |
| 17 | qwen3-5-27b | 0.807 |
| 18 | mimo-v2-flash | 0.806 |
| 19 | qwen3-6-35b-a3b | 0.804 |
| 20 | gemma-4-31b | 0.8 |
| 21 | qwen3-5-122b-a10b | 0.789 |
| 22 | gemma-4-26b-a4b | 0.771 |
| 23 | maple-preview | 0.751 |
| 24 | qwen3-5-35b-a3b | 0.746 |
| 25 | qwen3-235b-a22b-thinking-2507 | 0.741 |
LiveCodeBench v6
1 phaseActive
6th generation of the contamination-free competitive coding benchmark — continuously updated with fresh LeetCode, AtCoder, and Codeforces problems. Metric: pass@1.
Overview
LiveCodeBench v6
Quick answer: LiveCodeBench v6 is the sixth iteration of the contamination-resistant competitive programming benchmark by Jain et al. (2024). It collects fresh problems from LeetCode, AtCoder, and Codeforces, evaluating models on code generation, self-repair, code execution, and test output prediction. Qwen3.7 Max leads with 91.6% as of August 2026.
What Does LiveCodeBench v6 Test?
LiveCodeBench v6 evaluates large language models on four real-world coding scenarios using problems sourced continuously from competitive programming platforms. Because problems post-date training cutoffs, the benchmark resists contamination and memorisation effects that inflate scores on static benchmarks.
| Scenario | What it tests |
|---|---|
| Code generation | Write a correct solution from scratch |
| Self-repair | Fix a broken solution given test feedback |
| Code execution | Predict output for a given program |
| Test output prediction | Generate expected outputs for test cases |
How Is LiveCodeBench v6 Scored?
The primary metric is pass@1 — the probability that a single model-generated solution passes all hidden test cases. Scores range from 0 to 1 (reported as a percentage). Problems are weighted equally regardless of difficulty level.
LiveCodeBench v6 vs. LiveCodeBench (Earlier Versions)
Each version of LiveCodeBench incorporates a newer set of problems collected after the previous version's cutoff date. v6 reflects performance on problems released in the most recent window, making it the most up-to-date contamination-free coding signal available.
Key Facts
| Property | Value |
|---|---|
| Published | March 2024 |
| Problems | Continuously updated |
| Languages | Python (primary) |
| Metric | pass@1 |
| Score range | 0–1 |
| Top model | Qwen3.7 Max (0.916) |
FAQ
What is LiveCodeBench v6? LiveCodeBench v6 is the sixth major problem set released by the LiveCodeBench project. It evaluates AI coding ability using fresh competitive programming problems from LeetCode, AtCoder, and Codeforces that were published after model training cutoffs.
How is LiveCodeBench v6 different from HumanEval or MBPP? Unlike HumanEval and MBPP, which are static datasets that models may have memorised during training, LiveCodeBench collects new problems continuously to ensure contamination-free evaluation.
Who created LiveCodeBench? LiveCodeBench was created by Naman Jain, King Han, Alex Gu, Wen-Ding Li, and colleagues, and first described in the paper "LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code" (arXiv 2403.07974).
What score does the best model achieve on LiveCodeBench v6? Qwen3.7 Max currently achieves 0.916 (91.6%), ranking first among 53 evaluated models.