PerfCodeBench - Correct-and-Runnable 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 correct-and-runnable rate: the percentage of tasks whose generated code compiles, runs and passes the task oracle; higher is better.

Metric: Correct-and-runnable rate (%): share of tasks whose generated code compiles, runs and passes the oracle (compilation errors, runtime errors and wrong output all count as failures), 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-5.471.25#135
2Claude Opus 4.570.55#160
3Claude Sonnet 4.569.36#261
4GPT-561.81#178
5DeepSeek V4 Pro54.48#196
6Gemini 3.1 Flash51.83#312
7Seed 2.0 Lite47.57#300
8Qwen 3.6 Max47.25#193
9DeepSeek V4 Flash47.09#267
10Qwen 3.6 Plus45.2#238
11Gemma 4 26B A4B45.15#590
12Gemma 4 31B42.45#382
13Kimi K236.03#443
14Qwen 3.6 35B A3B32.09#481
15Llama 4 Maverick29.4#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-correct-and-runnable-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.