Skip to content

Benchmark

OSWorld 2

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages.

Models measured
16
Scores recorded
29
Sources
2
IQ weight
2.09

Score distribution

35.5%
Mean
27.3%
Median
24.3%
Std dev
8.3%
10th percentile
70.5%
90th percentile
2.8%
Min
2.8%77.9%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.09 for chat models, measured over 16 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude Fable 5.1· Steel.dev Agent Leaderboards data77.9%
2GPT-6 Astra· Steel.dev Agent Leaderboards data72.6%
3Claude Opus 5· Steel.dev Agent Leaderboards data70.6%
4GPT-6.1 Sol· Steel.dev Agent Leaderboards data70.5%
5Gemini 4 Argon· Steel.dev Agent Leaderboards data69.2%
6Muse Spark 1.3· Steel.dev Agent Leaderboards data66.9%
7Claude Fable 5· Steel.dev Agent Leaderboards data66.1%
8GPT-5.6 Sol· Steel.dev Agent Leaderboards data62.6%
9Qwen 3.8 Flash Next· Steel.dev Agent Leaderboards data52.3%
10Claude Opus 4.8· Steel.dev Agent Leaderboards data49.3%
11Claude Opus 4.7· Steel.dev Agent Leaderboards data49.1%
12Claude Opus 5max / batch tool· Epoch AI Benchmarking Hub data31.4%
13Claude Opus 5xhigh / batch tool· Epoch AI Benchmarking Hub data30.2%
14Claude Opus 5high / batch tool· Epoch AI Benchmarking Hub data29%
15GPT-5.6 Solmax / batch tool· Epoch AI Benchmarking Hub data27.3%
16Claude Opus 5medium / batch tool· Epoch AI Benchmarking Hub data25.3%
17Claude Opus 5low / batch tool· Epoch AI Benchmarking Hub data22.3%
18MiniMax M3· Steel.dev Agent Leaderboards data22.3%
19Qwen3.7 Plus· Steel.dev Agent Leaderboards data21.5%
20Claude Opus 4.8max / batched tool· Epoch AI Benchmarking Hub data20.6%
21Claude Opus 4.8max / standard· Epoch AI Benchmarking Hub data18.5%
22Claude Opus 4.7max / batched tool· Epoch AI Benchmarking Hub data18.2%
23Claude Opus 4.7max / standard· Epoch AI Benchmarking Hub data13.9%
24GPT 5.5xhigh / batch tool· Epoch AI Benchmarking Hub data13%
25Claude Sonnet 4.6medium / standard· Epoch AI Benchmarking Hub data9.3%
26Claude Sonnet 4.6max / standard· Epoch AI Benchmarking Hub data8.3%
27Kimi K2.6enabled / standard· Epoch AI Benchmarking Hub data4.6%
28MiniMax M3enabled / standard· Epoch AI Benchmarking Hub data4.6%
29Qwen3.7 Plusthinking / standard· Epoch AI Benchmarking Hub data2.8%

Data credits: Steel.dev Agent Leaderboards, MIT (licence) · Epoch AI Benchmarking Hub, CC BY 4.0 (licence)