PerfCodeBench - Reference-Beating Rate: leaderboard
Metric: Reference-beating rate (%): share of comparable tasks where the correct code matches or beats the reference implementation, 1,854 system-level performance-optimization tasks across C, C++, Go, Java, Python and CUDA; the model returns one replacement source file from the task metadata, interface contract and baseline in one shot; a candidate earns performance credit only if it compiles, runs and passes the task oracle; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 61.6 |
| 2 | Kimi K2.6 | 52.13 |
| 3 | GPT-5.4 | 51.82 |
| 4 | DeepSeek V4 Pro | 47.95 |
| 5 | Claude Opus 4.5 | 43.23 |
| 6 | DeepSeek V4 Flash | 43.05 |
| 7 | Seed 2.0 Lite | 40.68 |
| 8 | Gemini 3.1 Pro (Preview) | 40.09 |
| 9 | Qwen 3.6 27B | 39.6 |
| 10 | Qwen 3.6 Max | 37.19 |
| 11 | Claude Sonnet 4.5 | 32.81 |
| 12 | Qwen 3.6 Plus | 29.7 |
| 13 | Kimi K2 | 27.58 |
| 14 | Qwen 3.6 35B A3B | 22.51 |
| 15 | Llama 4 Maverick | 20.3 |
Interactive version: theaggregate.ai/benchmark?slug=perfcodebench-reference-beating-rate · How It Works · Data refreshed daily, snapshot 2026-10-07.