Unverified50% confidenceFactExact time
OSWorld 2.0采用细粒度检查点部分奖励评分，平均每个任务有27.25个检查点，不要求固定顺序
1
Sources
50%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
OSWorld 2.0测评：最强AI Agent完成率仅20%，长流程是致命短板
bilibili熊二等兵7/1/2026
Related Claims
UnverifiedIn practical tests, completing a search task using AnySearch Skill consumed 18 quota points, while not using it consumed 23 quota points, saving approximately 27%.64% similarUnverifiedThe tester estimated that Codex can complete approximately 30% of the tasks on the kanban board.57% similarUnverifiedMiniMax M2.7 asks 0.6 questions per task, but its clarification advancement rate is only 60.7%, indicating a problem of low-information-gain questioning56% similarUnverified根据Artificial Analysis数据，Fable完成单个任务约用33k token，Opus 5约用37k token55% similarUnverifiedMiniMax M3 scored an average of 58.3 points across 7 high-difficulty task tests.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/338172API
curl https://kongchang.com/api/v1/knowledge/claims/338172MCP
get_claim(id=338172)