korshunov.ai Everything that matters in ML
EN
Benchmark · agentic

OSWorld 2.0

9 results 7 models

2026 long-horizon re-cut of OSWorld: 108 multi-app desktop workflows, median human task ~1.6h. Not on the same scale as OSWorld / OSWorld-Verified — file here only when the quote says 2.0.

0 19.5 39 58.5 78 2026-07-28 2026-08-17 2026-09-05 Claude Opus 5 · 70 · 2026-07-28 Qwen-CUA · 18.5 · 2026-08-04 Qwen-CUA · 18.5 · 2026-08-04 Qwen-CUA-Max · 21.2 · 2026-08-04 Qwen-CUA-Max · 21.2 · 2026-08-04 Claude Fable 5.1 · 41.7 · 2026-09-01 Muse Spark 1.3 · 66.9 · 2026-09-04 GPT-6 Astra · 72.6 · 2026-09-04 GPT‑6 Astra · 72.6 · 2026-09-05
Claude Opus 5 Qwen-CUA Qwen-CUA-Max Claude Fable 5.1 Muse Spark 1.3 GPT-6 Astra GPT‑6 Astra
Timeline
Date Model Score Source
2026-09-05 GPT‑6 Astra 72.6% OpenAI releases GPT-6 Astra with improved computer use, coding, and cybersecurity capabilities
2026-09-04 GPT-6 Astra 72.6% OpenAI releases GPT-6 Astra, a computer-use model with 1.05M context
2026-09-04 Muse Spark 1.3 66.9% Meta releases Muse Spark 1.3 with fewer tool calls and tokens
2026-09-01 Claude Fable 5.1 41.7% Anthropic releases Claude Fable 5.1 with 52.6% on Terminal-Bench-Science and cheaper cache reads
2026-08-04 Qwen-CUA 18.5% Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent
2026-08-04 Qwen-CUA-Max 21.2% Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent
2026-08-04 Qwen-CUA 18.5% Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent
2026-08-04 Qwen-CUA-Max 21.2% Qwen releases Qwen-CUA, a 397B-A17B native computer-use agent
2026-07-28 Claude Opus 5 70.0% Anthropic releases Claude Opus 5 as a cost-effective alternative to Fable