PerfCodeBench - Faster-than-Baseline 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 tasks where the model returns correct code that runs faster than the baseline implementation; higher is better.

Metric: Faster-than-baseline rate (%): share of tasks where the model produces correct code that runs faster than the baseline 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-5.466.72#135
2Claude Opus 4.565.75#160
3Claude Sonnet 4.562.08#261
4GPT-559.55#178
5DeepSeek V4 Pro51.13#196
6Gemini 3.1 Flash46.98#312
7DeepSeek V4 Flash44.82#267
8Seed 2.0 Lite43.8#300
9Qwen 3.6 Max42.5#193
10Qwen 3.6 Plus39.75#238
11Gemma 4 26B A4B39.64#590
12Gemma 4 31B37.65#382
13Kimi K232.9#443
14Qwen 3.6 35B A3B27.94#481
15Llama 4 Maverick24.38#1110

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

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