PerfCodeBench - Efficiency Gap Closed: 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 whose correct code closes at least 80 percent of the baseline-to-reference performance gap; higher is better.

Metric: Share of comparable tasks (%) whose correct code closes at least 80 percent of the baseline-to-reference performance gap (CGRE at least 0.8), 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-572.91#178
2GPT-5.464.26#135
3Claude Opus 4.563.17#160
4Qwen 3.6 Max63.1#193
5DeepSeek V4 Pro62.88#196
6Kimi K2.661.7#192
7Qwen 3.6 27B61.39#351
8Gemini 3.1 Pro (Preview)60.67#99
9DeepSeek V4 Flash60.29#267
10Claude Sonnet 4.559.15#261
11Seed 2.0 Lite58.99#300
12Qwen 3.6 Plus54.85#238
13Gemini 3.1 Flash43.92#312
14Kimi K243.58#443
15Qwen 3.6 35B A3B41.26#481

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

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