Agentic computer use
OSWorld-Verified
Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Higher is better.
Rankings
Higher is better#186.1%
Qwen3.8-Max
285%
Claude Fable 5
285%
Claude Mythos 5via BenchLM
484.3%
Qwen3.8-27BOWvia BenchLM
583.4%
Claude Opus 4.8
683%
Gemini 3.6 Flash
781.2%
Claude Sonnet 5
880.8%
Muse Spark 1.1
978.7%
GPT-5.5
1078.4%
Gemini 3.5 Flash
1178%
Claude Opus 4.7
1276.2%
Gemini 3.1 Pro
1375%
GPT-5.4
1474%
Gemini 3.5 Flash-Lite
1573.3%
Qwen3.7-Plusvia BenchLM
1673.1%
Kimi K2.6via BenchLM
1772.7%
Claude Opus 4.6via BenchLM
1872.5%
Claude Sonnet 4.6
1972.1%
GPT-5.4 minivia BenchLM
2066.3%
Claude Opus 4.5via BenchLM
2165.9%
Muse GlimmerOW
2265.1%
Gemini 3.0 Flash
2364.7%
GPT-5.3-Codexvia BenchLM
2462.2%
Qwen3.5OW
2561.4%
Claude Sonnet 4.5via BenchLM
2653.3%
Muse Spark
2747.3%
GPT-5.2via BenchLM
2839%
GPT-5.4 nanovia BenchLM
Scores marked “via” above are quoted with attribution from BenchLM, retrieved 31 August 2026.
OSWorld-Verified — frequently asked questions
Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Higher is better.