New OSWorld leader: 79.38% success at $14.34/task, 40% cheaper than Claude
xwang_lk · x · 2026-10-06
Team results verified by OSWorld: Sai's agent now ranks #1 in partial success rate (79.38%) on real-world desktop tasks. The standout is efficiency — $14.34 per task vs. $24.11 for Anthropic's Claude Computer Use with Opus 5, achieving higher completion while cutting per-task inference cost by over 40%. The authors stress that reliably driving an OS hinges on the orchestration layer, grounding, and interaction primitives rather than raw compute.
More from coding & agent
- Gemini CLI PR adds custom OTLP headers to telemetry, enabling auth'd exports to Grafana and Datadog — jesussamuel-byte · 2026-10-06
- A 4-step Opus 5.5 + Hermes workflow that turns competitor pages and reviews into testable positioning angles — VibeMarketer_ · 2026-10-06
- Jeff Weinstein courts SaaS platforms to embed agent-driven purchasing flows with approval gates — jeff_weinstein · 2026-10-06
- Is Product the new DevRel? DevRel teams are merging into product and PM roles — gethackteam · 2026-10-06
- How 4 humans coordinate Claude Code, Codex, Muse and Hermes via Microsoft Planner kanban — DJAI9LAB · 2026-10-06
- HyperFrames Studio launches: a video editor built for agents, co-editing on one timeline — toolstelegraph · 2026-10-06