Claude Opus 5 tops OSWorld with 70.6% partial-credit, about 30% strict success
ysu_nlp · x · 2026-07-27
Claude’s team is being congratulated for a major jump on OSWorld, but the post notes that the headline 70.6% is partial-credit scoring. Under strict binary task-level success, Opus 5 is still only about 30%, and the team says it is working on improving task quality so the benchmark is more robust on hard tasks.
More from Models
- Opus 5 looks stronger than Opus 4.8, but picks up some Fable quirks — eyishazyer · 2026-07-27
- Opus 5 can build a full custom first-person shooter from one prompt — eyishazyer · 2026-07-27
- Three days after launch, Opus 5 is already making consultant-grade decks — eyishazyer · 2026-07-27
- Writer says ChatGPT still trails Claude in creative writing after a month of testing — NeonXEExperiment · 2026-07-27
- Anthropic publishes an Opus 5 prompt guide built around subtraction, not addition — xiaohu · 2026-07-27
- Google says 86% of Gemini chats in a 14.65M-conversation study were non-work related — xiaohu · 2026-07-27