Agentic · tools

OSWorld 2.1 Offline 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 Offline is 87.2% by Claude Opus 5.5. 69 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	Muse Wonder Full	88.1%	estimated ± 7.9 pp, low confidence
2	Claude Opus 5.5	87.2%	measured
3	Kimi K3.1 Pro	87.1%	estimated ± 7.9 pp, low confidence
4	Claude Sonnet 5.5	83.9%	measured
5	Gemini 4 Carbon	81.5%	measured
6	Muse Wonder	81.0%	estimated ± 7.9 pp, low confidence
7	Claude Fable 5.1	79.7%	estimated ± 7.9 pp, low confidence
8	Claude Opus 5	77.1%	estimated ± 7.9 pp, low confidence
9	Grok 4.7	76.5%	estimated ± 7.9 pp, low confidence
10	Gemini 4 Argon	74.7%	measured
11	GPT-6.1 Terra	74.2%	measured
12	MiMo-V2.6-Pro	74.1%	estimated ± 7.9 pp, low confidence
13	Muse Spark 1.3	73.7%	estimated ± 7.9 pp, low confidence
14	Qwen3.8 Max Preview	72.6%	estimated ± 7.9 pp, low confidence
15	Claude Haiku 5.5	72.4%	measured
16	GLM-5.3	71.3%	estimated ± 7.9 pp, low confidence
17	Qwen3.8-Flash-Next	69.6%	estimated ± 7.9 pp, low confidence
18	Grok 4.6	68.7%	estimated ± 7.9 pp, low confidence
19	Ling 3.1 Flash	68.7%	estimated ± 7.9 pp, low confidence
20	Claude Fable 5	68.1%	estimated ± 7.9 pp, low confidence
21	GPT-5.6 Sol	67.9%	estimated ± 7.9 pp, low confidence
22	MiMo-V2.6-Flash	67.9%	estimated ± 7.9 pp, low confidence
23	DeepSeek V4.1 Flash 1017	67.3%	estimated ± 7.9 pp, low confidence
24	GPT-6.1 Sol	67.0%	estimated ± 7.9 pp, low confidence
25	DeepSeek V4 Pro 0813	66.5%	estimated ± 7.9 pp, low confidence
26	Step 5 Preview	66.2%	estimated ± 7.9 pp, low confidence
27	GPT-6 Astra	65.2%	estimated ± 7.9 pp, low confidence
28	Kimi K3	61.8%	estimated ± 7.9 pp, low confidence
29	GPT-6 Sol	59.9%	estimated ± 7.9 pp, low confidence
30	Muse Spark 1.2	57.0%	estimated ± 7.9 pp, low confidence
31	Claude Sonnet 5	56.3%	estimated ± 7.9 pp, low confidence
32	GPT-5.6 Luna	56.0%	estimated ± 7.9 pp, low confidence
33	MiniMax M3.1 Flash	55.9%	estimated ± 7.9 pp, low confidence
34	Claude Opus 4.8	55.5%	estimated ± 7.9 pp, low confidence
35	GPT-5.6 Terra	55.4%	estimated ± 7.9 pp, low confidence
36	Gemini 3.8 Flash	54.1%	estimated ± 7.9 pp, low confidence
37	DeepSeek V4 Flash 0731	53.9%	estimated ± 7.9 pp, low confidence
38	Mistral Large 4	53.0%	estimated ± 7.9 pp, low confidence
39	Qwen3.8-27B	52.6%	estimated ± 7.9 pp, low confidence
40	Muse Spark Open	51.3%	estimated ± 7.9 pp, low confidence
41	Gemini 3.7 Flash	50.5%	estimated ± 7.9 pp, low confidence
42	Grok 4.5	49.9%	estimated ± 7.9 pp, low confidence
43	GLM-5.2	48.9%	estimated ± 7.9 pp, low confidence
44	GPT-6 Luna	48.9%	measured
45	Gemini 3.1 Pro	48.5%	estimated ± 9.2 pp, low confidence
46	Claude Opus 4.7 (Adaptive)	47.3%	estimated ± 7.9 pp, low confidence
47	GPT-5.5	47.2%	estimated ± 7.9 pp, low confidence
48	Gemini 3.5 Flash	46.7%	estimated ± 7.9 pp, low confidence
49	Gemini 3.6 Flash	41.8%	estimated ± 7.9 pp, low confidence
50	GPT-5.4	38.8%	estimated ± 7.9 pp, low confidence
51	MiniMax M3	38.5%	estimated ± 7.9 pp, low confidence
52	Hy3 Preview	36.4%	estimated ± 7.9 pp, low confidence
53	Muse Spark 1.1	36.0%	estimated ± 7.9 pp, low confidence
54	Apodex 1.1	34.1%	estimated ± 7.9 pp, low confidence
55	Apodex 1.1 Mini	34.1%	estimated ± 7.9 pp, low confidence
56	Quasar 438B	33.6%	estimated ± 7.9 pp, low confidence
57	Beam 1.0 501B-A23B	33.3%	estimated ± 7.9 pp, low confidence
58	Nemotron 3.5 Super VL 120B A12B	32.0%	estimated ± 7.9 pp, low confidence
59	Nemotron 3.5 Super VL 120B A12B NVFP4	32.0%	estimated ± 7.9 pp, low confidence
60	Nemotron 3 Ultra	32.0%	estimated ± 7.9 pp, low confidence
61	Ling 3.0 Flash VL	31.8%	estimated ± 7.9 pp, low confidence
62	Qwen3.7 Max	29.3%	estimated ± 7.9 pp, low confidence
63	Inkling-Small	28.9%	estimated ± 7.9 pp, low confidence
64	MiMo-V2.5-Pro	28.6%	estimated ± 7.9 pp, low confidence
65	GLM-5.1	28.3%	estimated ± 7.9 pp, low confidence
66	Solar Pro 4	27.5%	estimated ± 7.9 pp, low confidence
67	Grok 4.3	25.7%	estimated ± 7.9 pp, low confidence
68	Inkling	24.8%	estimated ± 7.9 pp, low confidence
69	Hy3	24.1%	estimated ± 7.9 pp, low confidence
70	Kimi K2.6	22.0%	estimated ± 7.9 pp, low confidence
71	Kimi K2.7 Code	22.0%	estimated ± 7.9 pp, low confidence
72	Step 3.7 Flash	20.2%	estimated ± 7.9 pp, low confidence
73	GLM-4.7	20.1%	estimated ± 7.9 pp, low confidence
74	GPT-5.4 mini	19.9%	estimated ± 7.9 pp, low confidence
75	MiniMax M2.7	19.8%	estimated ± 7.9 pp, low confidence
76	Muse Spark	18.8%	estimated ± 7.9 pp, low confidence
77	Claude Haiku 4.5	15.7%	measured

OSWorld 2.1 Offline leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General