PerfCodeBench - Reference-Beating Rate: leaderboard

PerfCodeBench (2026) tests whether LLMs can optimize real system-level code: 1,854 executable tasks over C, C++, Go, Java, Python and CUDA, where the model rewrites a baseline and is timed against a baseline and an expert reference. This board is the percentage of comparable tasks where the correct code matches or beats the expert reference implementation; higher is better.

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 February 2027. 20 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-561.6#178
2Kimi K2.652.13#192
3GPT-5.451.82#135
4DeepSeek V4 Pro47.95#196
5Claude Opus 4.543.23#160
6DeepSeek V4 Flash43.05#267
7Seed 2.0 Lite40.68#300
8Gemini 3.1 Pro (Preview)40.09#99
9Qwen 3.6 27B39.6#351
10Qwen 3.6 Max37.19#193
11Claude Sonnet 4.532.81#261
12Qwen 3.6 Plus29.7#238
13Kimi K227.58#443
14Qwen 3.6 35B A3B22.51#481
15Gemini 3.1 Flash22.02#312

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=perfcodebench-reference-beating-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.