Skip to content
Weights Not Walls
Menu

OSWorld 2.0

Reported

No description yet: we only describe an eval set when we can do it correctly.

Kind
Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
Link
https://osworld-v2.xlang.ai/
Score shown
Epoch's column "Binary accuracy", as a percentage, higher is better; random guessing scores about 0.0%. The source gives no error bars.
Results
16 in all: 2 for open-weight models, 14 for closed ones. 2 linked to a model page here.

Best open-weight result: moonshotai/Kimi-K2.6, 4.6%, number 14 of 16 results.

OSWorld 2.0 results
#ModelWeightsSettingScoreDateSource
1claude-opus-5_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: batch tool; Step budget: 50031.4%not datedosworld-v2.xlang.ai
2claude-opus-5_xhighAnthropicclosedeffort xhigh; Reasoning: xhigh; Tool setting: batch tool; Step budget: 50030.2%not datedosworld-v2.xlang.ai
3claude-opus-5_highAnthropicclosedeffort high; Reasoning: high; Tool setting: batch tool; Step budget: 50029.0%not datedosworld-v2.xlang.ai
4gpt-5.6-sol_maxOpenAIclosedeffort max; Reasoning: max; Tool setting: batch tool; Step budget: 50027.3%not datedosworld-v2.xlang.ai
5claude-opus-5_mediumAnthropicclosedeffort medium; Reasoning: medium; Tool setting: batch tool; Step budget: 50025.3%not datedosworld-v2.xlang.ai
6claude-opus-5_lowAnthropicclosedeffort low; Reasoning: low; Tool setting: batch tool; Step budget: 50022.3%not datedosworld-v2.xlang.ai
7claude-opus-4-8_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: batched tool; Step budget: 50020.6%not datedosworld-v2.xlang.ai
8claude-opus-4-8_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: standard; Step budget: 50018.5%not datedosworld-v2.xlang.ai
9claude-opus-4-7_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: batched tool; Step budget: 50018.2%not datedosworld-v2.xlang.ai
10claude-opus-4-7_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: standard; Step budget: 50013.9%not datedosworld-v2.xlang.ai
11gpt-5.5_xhighOpenAIclosedeffort xhigh; Reasoning: xhigh; Tool setting: batch tool; Step budget: 50013.0%not datedosworld-v2.xlang.ai
12claude-sonnet-4-6_mediumAnthropicclosedeffort medium; Reasoning: medium; Tool setting: standard; Step budget: 5009.3%not datedosworld-v2.xlang.ai
13claude-sonnet-4-6_maxAnthropicclosedeffort max; Reasoning: max; Tool setting: standard; Step budget: 5008.3%not datedosworld-v2.xlang.ai
14moonshotai/Kimi-K2.6Epoch: kimi-k2.6, MoonshotopenReasoning: enabled; Tool setting: standard; Step budget: 5004.6%not datedosworld-v2.xlang.ai
15MiniMaxAI/MiniMax-M3Epoch: MiniMax-M3, MiniMaxopenReasoning: enabled; Tool setting: standard; Step budget: 5004.6%not datedosworld-v2.xlang.ai
16qwen3.7-plusAlibabaclosedReasoning: thinking; Tool setting: standard; Step budget: 5002.8%not datedosworld-v2.xlang.ai

Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method