Benchmarks · coding
GSO software optimization, Opt@1 (UC Berkeley)
An agent gets a real codebase and a performance test and must make the code as fast as an expert developer's optimization; 102 tasks across 10 codebases and 5 languages. Maintained by GSO team. Score: Opt@1: share of tasks where a single attempt reaches at least 95 % of the human speedup and passes the correctness tests.
26 models · 28 results · data as of 2026-10-29 · gso-bench.github.io ↗
Leaderboard
Opus 4.8 xhigh
All results
28 rows| # | Model | Effort | Score | Harness |
|---|---|---|---|---|
| 1 | Opus 4.8 | xhigh | 47.1% | OpenHands |
| 2 | Opus 4.7 | high | 44.1% | OpenHands |
| 3 | Opus 4.6 | high | 41.2% | OpenHands |
| 4 | GPT-5.5 | xhigh | 40.2% | OpenHands |
| 5 | Sonnet 5 | xhigh | 37.3% | OpenHands |
| 6 | Opus 4.6 | max | 33.3% | OpenHands |
| 7 | GPT-5.4 | xhigh | 31.4% | OpenHands |
| 8 | GPT-5.2 | high | 27.4% | OpenHands |
| 9 | Opus 4.5 | high | 26.5% | OpenHands |
| 10 | GPT-5.4 | high | 25.5% | OpenHands |
| 11 | Gemini 3.1 Pro | high | 22.6% | OpenHands |
| 12 | Gemini 3 Pro | high | 18.6% | OpenHands |
| 13 | Sonnet 4.5 | high | 14.7% | OpenHands |
| 14 | GPT-5.1 | high | 13.7% | OpenHands |
| 15 | Gemini 3 Flash | high | 9.8% | OpenHands |
| 16 | o3 | high | 8.8% | OpenHands |
| 17 | Opus 4 | medium | 6.9% | OpenHands |
| 18 | GPT-5 | high | 6.9% | OpenHands |
| 19 | Sonnet 4 | high | 4.9% | OpenHands |
| 20 | Kimi K2 | medium | 4.9% | OpenHands |
| 21 | Qwen 3 Coder | medium | 4.9% | OpenHands |
| 22 | Sonnet 3.5 | medium | 4.6% | OpenHands |
| 23 | Gemini 2.5 Pro | high | 3.9% | OpenHands |
| 24 | Sonnet 3.7 | medium | 3.8% | OpenHands |
| 25 | o4 mini | high | 3.6% | OpenHands |
| 26 | GLM 4.5 Air | 2.9% | OpenHands | |
| 27 | o3 mini | high | 1.3% | OpenHands |
| 28 | GPT-4o | high | 0.0% | OpenHands |