Skip to content

Benchmark

OSWorld 2.0 · Partial

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

Models measured
24
Scores recorded
76
Sources
1
IQ weight
No weight published

Score distribution

59.3%
Mean
62.2%
Median
16.4%
Std dev
41.4%
10th percentile
75.4%
90th percentile
8.3%
Min
8.3%87.2%
The band covers the middle 80% of scores. The line marks the median.

Leaderboard

#ModelScore
51Claude Opus 4.8max · 108 tasks · GUI-only · no exec · 500 steps · partial reward· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54.8%
52GPT-6 SolOSWorld 2.0 offline set · v2026.08.08 · partial reward · OpenAI release chart · medium reasoning · displayed cost $1.38/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54%
53GPT-6 SolOSWorld 2.0 offline v2026.08.08 · medium effort · $1.3169 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.54%
54GPT-5.6 LunaOSWorld 2.0 offline v2026.08.08 · max effort · $0.4909 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.52.7%
55GPT-6 LunaOSWorld 2.0 offline v2026.08.08 · max effort · $0.2678 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.52.7%
56Qwen 3.8 Flash NextQwen3.8-Flash-Next release/model card · aggregate partial reward· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.52.3%
57GPT-5.6 SolOSWorld 2.0 offline v2026.08.08 · medium effort · $2.7275 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.49.7%
58GPT-6 Lunaadaptive thinking · max effort · default sampling · 5-trial mean · context ≤1M unless noted · 82-task offline subset · OpenAI API · max · context compaction· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.48.9%
59GPT-5.6 LunaOSWorld 2.0 offline v2026.08.08 · xhigh effort · $0.3573 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.48%
60Gemini 3.7 Flashpartial · max over 3 runs· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.47.9%
61Muse Spark 1.2xhigh · OSWorld 2.0 v06.24 · 108 tasks · common internal framework · partial reward· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.47.6%
62GPT 5.5xhigh · 108 tasks · GUI-only · no exec · 500 steps · partial reward· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.47.5%
63Muse Spark 1.1Meta Model API · xhigh · 108 tasks · GUI-only · no exec · 500 steps · partial reward· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.47.3%
64GPT-6 LunaOSWorld 2.0 offline v2026.08.08 · xhigh effort · $0.1639 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.46.7%
65Qwen3.8 MaxQwen3.8-Max release vision/video table · aggregate partial reward· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.46.7%
66GPT-6 SolOSWorld 2.0 offline set · v2026.08.08 · partial reward · OpenAI release chart · low reasoning · displayed cost $1.01/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.43.9%
67GPT-6 SolOSWorld 2.0 offline v2026.08.08 · low effort · $0.9675 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.43.9%
68GPT-6 LunaOSWorld 2.0 offline v2026.08.08 · high effort · $0.1227 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.41.4%
69GPT-5.6 LunaOSWorld 2.0 offline v2026.08.08 · high effort · $0.1614 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.35.9%
70GPT-6 LunaOSWorld 2.0 offline v2026.08.08 · medium effort · $0.0624 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.31.5%
71Gemini 3.1 Pro Previewhigh · 108 tasks · GUI-only · no exec · 500 steps · partial reward· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.30.6%
72GPT-5.6 SolOSWorld 2.0 offline v2026.08.08 · low effort · $0.9086 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.29.8%
73GPT-5.6 LunaOSWorld 2.0 offline v2026.08.08 · medium effort · $0.0516 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.22.7%
74Claude Haiku 4.5adaptive thinking · max effort · default sampling · 5-trial mean · context ≤1M unless noted · 82-task offline subset · fixed thinking budget · 5 runs· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.15.7%
75GPT-5.6 LunaOSWorld 2.0 offline v2026.08.08 · low effort · $0.0185 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.12%
76GPT-6 LunaOSWorld 2.0 offline v2026.08.08 · low effort · $0.0297 average cost/task· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.8.3%

Showing 51–76 of 76 results.

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.