Why this benchmark is useful
When you want contest-style coding with fresher problems that resist training-data leakage better than classic static suites.
LiveCodeBench
Contamination-resistant coding problems drawn from recent contest releases — AA reruns models on the public harness.
What this does not measure
- Multi-file refactors or IDE agent loops.
- Private codebase context — problems are public contest slices only.
Coverage map
- Task family
- Coding
- Format
- Timed coding tasks with public grader.
- Scoring
- pass@1 percentage from AA runs.
- Maintainer
- LiveCodeBench authors
How to read the scores
pass@1 from AA public harness runs. Timed contest slices, not multi-file IDE agents — do not treat as production codebase skill.
What it does not cover
- Multi-file refactors or IDE agent loops.
- Private codebase context — problems are public contest slices only.
Scores
31 rowsEvidence ledger
3131 rows
91.7%best score
31source-checked
1sources
2026-09-01source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- Gemini 3 Pro91.7% ·
- Gemini 3 Flash Preview (Reasoning)90.8% ·
- GLM-4.7 (Reasoning)89.4% ·
- GPT-5.2 (medium)89.4% ·
- GPT-5.288.9% ·
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | Gemini 3 ProGoogle | 2025-11-18 | 91.7% | 2026-09-01 | Source-checked |
| 2 | Gemini 3 Flash Preview (Reasoning)Google | 2025-12-17 | 90.8% | 2026-09-01 | Source-checked |
| 3 | GLM-4.7 (Reasoning)Zhipu | 2025-12-22 | 89.4% | 2026-09-01 | Source-checked |
| 4 | GPT-5.2 (medium)OpenAI | 2025-12-11 | 89.4% | 2026-09-01 | Source-checked |
| 5 | GPT-5.2OpenAI | 2025-12-11 | 88.9% | 2026-09-01 | Source-checked |
| 6 | Claude Opus 4.5 (Reasoning)Anthropic | 2025-11-24 | 87.1% | 2026-09-01 | Source-checked |
| 7 | GPT-5.1OpenAI | 2025-11-13 | 86.8% | 2026-09-01 | Source-checked |
| 8 | MiMo-V2-Flash (Reasoning)Xiaomi | 2025-12-16 | 86.8% | 2026-09-01 | Source-checked |
| 9 | DeepSeek V3.2 (Reasoning)DeepSeek | 2025-12-01 | 86.2% | 2026-09-01 | Source-checked |
| 10 | Gemini 3 Pro Preview (low)Google | 2025-11-18 | 85.7% | 2026-09-01 | Source-checked |
| 11 | Kimi K2 ThinkingMoonshot | 2025-11-06 | 85.3% | 2026-09-01 | Source-checked |
| 12 | GPT-5.1 Codex (high)OpenAI | 2025-11-13 | 84.9% | 2026-09-01 | Source-checked |
| 13 | GPT-5 (high)OpenAI | 2025-08-07 | 84.6% | 2026-09-01 | Source-checked |
| 14 | GPT-5 Codex (high)OpenAI | 2025-09-23 | 84% | 2026-09-01 | Source-checked |
| 15 | Grok 4xAI | 2025-07-10 | 81.9% | 2026-09-01 | Source-checked |
| 16 | MiniMax-M2.1MiniMax | 2025-12-23 | 81% | 2026-09-01 | Source-checked |
| 17 | OpenAI o3OpenAI | 2025-04-16 | 80.8% | 2026-09-01 | Source-checked |
| 18 | Gemini 3 Flash PreviewGoogle | 2025-12-17 | 79.7% | 2026-09-01 | Source-checked |
| 19 | DeepSeek R1DeepSeek | 2025-01-20 | 77% | 2026-09-01 | Source-checked |
| 20 | GPT-5 (low)OpenAI | 2025-08-07 | 76.3% | 2026-09-01 | Source-checked |
| 21 | Claude Opus 4.5Anthropic | 2025-11-24 | 73.8% | 2026-09-01 | Source-checked |
| 22 | Claude 4.5 Sonnet (Reasoning)Anthropic | 2025-09-29 | 71.4% | 2026-09-01 | Source-checked |
| 23 | GPT-5 (medium)OpenAI | 2025-08-07 | 70.3% | 2026-09-01 | Source-checked |
| 24 | OpenAI o1OpenAI | 2024-09-12 | 67.9% | 2026-09-01 | Source-checked |
| 25 | Claude 4.1 Opus (Reasoning)Anthropic | 2025-08-05 | 65.4% | 2026-09-01 | Source-checked |
| 26 | Claude 4 Opus (Reasoning)Anthropic | 2025-05-22 | 63.6% | 2026-09-01 | Source-checked |
| 27 | Mistral Large 3Mistral | 2025-12-02 | 46.5% | 2026-09-01 | Source-checked |
| 28 | Llama 4 MaverickMeta | 2026-04-05 | 39.7% | 2026-09-01 | Source-checked |
| 29 | Llama 4 ScoutMeta | 2026-04-05 | 29.9% | 2026-09-01 | Source-checked |
| 30 | Command ACohere | 2026-03-15 | 28.7% | 2026-09-01 | Source-checked |
| 31 | Claude 3 OpusAnthropic | 2024-03-04 | 27.9% | 2026-09-01 | Source-checked |
What it covers
This benchmark sits in the coding family. It uses Timed coding tasks with public grader. The score should travel with its task format, scoring method, source date, and benchmark version.
Where it breaks down
No ceiling note is recorded yet. Treat clustering near the top as a warning that the benchmark may no longer separate frontier models.
Receipts
Sources and further reading
- AA LiveCodeBench· checked 2026-06-10
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does LiveCodeBench measure?
Contamination-resistant coding problems drawn from recent contest releases — AA reruns models on the public harness.
What does a high LiveCodeBench score not prove?
A strong LiveCodeBench result says nothing about:
- Multi-file refactors or IDE agent loops.
- Private codebase context — problems are public contest slices only.
How is LiveCodeBench scored?
pass@1 percentage from AA runs. Task format: Timed coding tasks with public grader.
Is LiveCodeBench saturated?
LiveCodeBench is currently marked Active in the atlas.
Can LiveCodeBench results be contaminated by training data?
Contamination risk for LiveCodeBench is graded Low contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.