Agentic · tools

OSWorld 2.1 leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

As of 2026-10-29, the highest measured score on OSWorld 2.1 is 81.8% by Claude Opus 5.5. 77 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	Claude Opus 5.5	81.8%	measured
2	Claude Fable 5.1	80.7%	measured
3	Claude Sonnet 5.5	80.1%	measured
4	Gemini 4 Carbon	79.9%	estimated ± 2.8 pp, low confidence
5	Gemini 4 Argon	79.9%	estimated ± 2.8 pp, low confidence
6	MiniMax M3.1 Flash	79.8%	estimated ± 2.8 pp, medium confidence
7	Grok 4.7	79.7%	estimated ± 2.8 pp, medium confidence
8	GPT-6.1 Sol	79.7%	estimated ± 2.8 pp, medium confidence
9	Kimi K3.1 Pro	79.6%	measured
10	GLM-5.3	79.6%	estimated ± 2.8 pp, medium confidence
11	GPT-6.1 Terra	79.5%	estimated ± 2.8 pp, medium confidence
12	Gemini 3.8 Flash	79.5%	estimated ± 2.8 pp, medium confidence
13	Mistral Large 4	79.5%	estimated ± 2.8 pp, medium confidence
14	MiMo-V2.6-Pro	79.4%	estimated ± 2.8 pp, medium confidence
15	GPT-6 Luna	78.9%	estimated ± 2.8 pp, medium confidence
16	Gemini 3.1 Pro	78.6%	estimated ± 7.8 pp, medium confidence
17	Step 5 Preview	78.5%	estimated ± 2.8 pp, medium confidence
18	Muse Wonder Full	78.4%	measured
19	Muse Spark Open	78.3%	estimated ± 2.8 pp, medium confidence
20	Beam 1.0 501B-A23B	78.0%	estimated ± 2.8 pp, medium confidence
21	Qwen3.8-27B	78.0%	estimated ± 2.8 pp, medium confidence
22	GPT-6 Astra	75.8%	measured
23	Claude Opus 5	74.0%	measured
24	Muse Spark 1.3	73.3%	estimated ± 6.0 pp, medium confidence
25	Qwen3.8 Max Preview	72.8%	estimated ± 6.0 pp, medium confidence
26	Muse Wonder	72.6%	measured
27	Qwen3.8-Flash-Next	71.3%	estimated ± 6.0 pp, medium confidence
28	Grok 4.6	70.9%	estimated ± 6.0 pp, medium confidence
29	Ling 3.1 Flash	70.9%	estimated ± 6.0 pp, medium confidence
30	Claude Fable 5	70.5%	estimated ± 6.0 pp, medium confidence
31	GPT-5.6 Sol	70.4%	estimated ± 6.0 pp, medium confidence
32	MiMo-V2.6-Flash	70.4%	estimated ± 6.0 pp, medium confidence
33	DeepSeek V4.1 Flash 1017	70.1%	estimated ± 6.0 pp, medium confidence
34	DeepSeek V4 Pro 0813	69.7%	estimated ± 6.0 pp, medium confidence
35	Claude Haiku 5.5	68.5%	estimated ± 2.8 pp, low confidence
36	Kimi K3	67.2%	estimated ± 6.0 pp, medium confidence
37	GPT-6 Sol	66.2%	estimated ± 6.0 pp, medium confidence
38	Claude Haiku 4.5	64.8%	estimated ± 7.8 pp, medium confidence
39	Claude Sonnet 4.6	64.8%	estimated ± 7.8 pp, medium confidence
40	GLM-5.3-Flash	64.8%	estimated ± 7.8 pp, medium confidence
41	GPT-5.3-Codex-Spark	64.8%	estimated ± 7.8 pp, medium confidence
42	Grok 4.20	64.8%	estimated ± 7.8 pp, medium confidence
43	Llama 4 Maverick	64.8%	estimated ± 7.8 pp, medium confidence
44	Qwen3.8 Max	64.8%	estimated ± 7.8 pp, medium confidence
45	Muse Spark 1.2	64.6%	estimated ± 6.0 pp, medium confidence
46	GPT-5.6 Luna	64.0%	estimated ± 6.0 pp, low confidence
47	Claude Opus 4.8	63.8%	estimated ± 6.0 pp, low confidence
48	GPT-5.6 Terra	63.7%	estimated ± 6.0 pp, low confidence
49	Ling 3.0 Flash Sante	63.6%	estimated ± 8.3 pp, low confidence
50	DeepSeek V4 Flash 0731	62.9%	estimated ± 6.0 pp, low confidence
51	Ornith-1.5-397B	62.0%	estimated ± 8.3 pp, low confidence
52	Hy4 preview	61.0%	estimated ± 8.3 pp, low confidence
53	Gemini 3.7 Flash	60.9%	estimated ± 6.0 pp, low confidence
54	Grok 4.5	60.5%	estimated ± 6.0 pp, low confidence
55	Sage-A1.5	60.5%	estimated ± 8.3 pp, low confidence
56	GLM-5.2	59.9%	estimated ± 6.0 pp, low confidence
57	Claude Opus 4.7 (Adaptive)	59.0%	estimated ± 6.0 pp, low confidence
58	GPT-5.5	58.9%	estimated ± 6.0 pp, low confidence
59	Gemini 3.5 Flash	58.6%	estimated ± 6.0 pp, low confidence
60	Claude Sonnet 5	57.0%	measured
61	dots3-note Preview	56.8%	estimated ± 8.3 pp, low confidence
62	Gemini 3.6 Flash	55.6%	estimated ± 6.0 pp, low confidence
63	Beam 1.0 501B-A23B FP8	54.6%	estimated ± 8.3 pp, low confidence
64	Beam 1.0 501B-A23B NVFP4	54.6%	estimated ± 8.3 pp, low confidence
65	GLM-5	53.6%	estimated ± 8.3 pp, low confidence
66	GPT-5.4	53.6%	estimated ± 6.0 pp, low confidence
67	MiniMax M3	53.4%	estimated ± 6.0 pp, low confidence
68	Hy3 Preview	52.0%	estimated ± 6.0 pp, low confidence
69	Muse Spark 1.1	51.8%	estimated ± 6.0 pp, low confidence
70	Apodex 1.1	50.5%	estimated ± 6.0 pp, low confidence
71	Apodex 1.1 Mini	50.5%	estimated ± 6.0 pp, low confidence
72	Quasar 438B	50.2%	estimated ± 6.0 pp, low confidence
73	Agents-A1	49.7%	estimated ± 8.3 pp, low confidence
74	Step 3.7 Flash	49.1%	estimated ± 8.3 pp, low confidence
75	Nemotron 3.5 Super VL 120B A12B	49.0%	estimated ± 6.0 pp, low confidence
76	Nemotron 3.5 Super VL 120B A12B NVFP4	49.0%	estimated ± 6.0 pp, low confidence
77	Nemotron 3 Ultra	49.0%	estimated ± 6.0 pp, low confidence
78	Ling 3.0 Flash VL	48.9%	estimated ± 6.0 pp, low confidence
79	Qwen3.7 Max	47.1%	estimated ± 6.0 pp, low confidence
80	Inkling-Small	46.9%	estimated ± 6.0 pp, low confidence
81	MiMo-V2.5-Pro	46.7%	estimated ± 6.0 pp, low confidence
82	GLM-5.1	46.4%	estimated ± 6.0 pp, low confidence
83	Solar Pro 4	45.8%	estimated ± 6.0 pp, low confidence
84	Grok 4.3	44.5%	estimated ± 6.0 pp, low confidence
85	Hy3	43.4%	estimated ± 6.0 pp, low confidence
86	Inkling	0.0%	estimated ± 2.8 pp, low confidence

OSWorld 2.1 leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General