Amodei guessed that by about November 2025, AI models would be able to perform computer-use tasks very, very reliably.

The claim is that by about November 2025 AI models will perform computer-use tasks (operating a GUI computer to complete tasks) 'very, very reliably', compared with a state of the art of about 15–22% on OSWorld in late 2024. Reliability is measured mainly by the best reported OSWorld (or OSWorld-Verified) score, with human-level performance (about 72%) as the benchmark for 'very reliable'. True: by 30 Nov 2025 the best reported score is 70% or higher, or comparable evidence shows near-human reliability on everyday computer tasks. Mostly true: 65% to under 70%. Mostly false: 40% to under 65% (large progress, but not very reliable). False: below 40%, or dated reports from late 2025 describing computer-use agents as still unreliable for routine tasks.

Set before the research began.

1. E1 Introducing Claude Sonnet 4.5 Anthropic · 29 September 2025 Own record

   “On OSWorld, a benchmark that tests AI models on real-world computer tasks, Sonnet 4.5 now leads at 61.4%. Just four months ago, Sonnet 4 held the lead at 42.2%.”

2. E2 Introducing OSWorld-Verified XLANG Lab (HKU), OSWorld benchmark maintainers · 28 July 2025

   “With human performance estimated at ~72% from our original study, the best current systems have now reached 84.4% of human capability (CoACT-1 at 60.76%).”

3. E3 osworld.md Steel (leaderboard.steel.dev) · 16 October 2026

   “| 13 | Claude Opus 4.5 | 66.3% | Anthropic | OSWorld-Verified self-reported result; reported on anthropic.com. | https://www.anthropic.com/news/claude-opus-4-5 | — |”

4. E4 OSWorld | AI Wiki AI Wiki (aiwiki.ai) · 7 May 2026

   “Within five months that number had been overtaken by frontier production systems, with Claude Opus 4.5 reaching 66.3% and Claude Opus 4.6 crossing the 72% human baseline in February 2026.”

5. E5 Introducing OSWorld-Verified XLANG Lab (xlang.ai) · 28 July 2025

   “With human performance estimated at ~72% from our original study, the best current systems have now reached 84.4% of human capability (CoACT-1 at 60.76%).”

6. E6 OSWorld | AI Wiki AI Wiki (aiwiki.ai) · 7 May 2026

   “| November 2025 | Claude Opus 4.5 | Anthropic | 66.3% | n/a | First model above 60% on main leaderboard |”

7. E7 OSWorld | AI Wiki AI Wiki (aiwiki.ai) · 7 May 2026

   “Third, the trajectory crossed the 72.36% human baseline in February 2026 when Claude Opus 4.6 reached 72.7%.”

8. E8 OSWorld | AI Wiki AI Wiki (aiwiki.ai) · 7 May 2026

   “The September 2025 jump to 61.4% in Claude Sonnet 4.5 was widely characterized in the developer press as the first "production-grade" computer-use Claude.”

9. E9 Claude Opus 4.6: Features, Benchmarks, and Pricing Guide Digital Applied · 5 February 2026

   “| OSWorld | 72.7% | 66.3% | — | — |”

10. E10 Introducing Claude Opus 4.5 Anthropic · 24 November 2025 Own record

    “Each of these updates takes advantage of Claude Opus 4.5’s market-leading performance in using computers, spreadsheets, and handling long-running tasks.”

11. E11 Anthropic releases Opus 4.5 with new Chrome and Excel integrations | TechCrunch TechCrunch · 24 November 2025

    “Anthropic also emphasized Opus’ computer use and spreadsheet capabilities, and launched a number of parallel products to showcase how the model holds up in those settings.”

An independent wiki timeline lists Claude Opus 4.5 at 66.3% on OSWorld in November 2025 (E6, E4), and Anthropic's launch post for that model is dated 2025-11-24 (E10). A third-party leaderboard also lists this 66.3% OSWorld-Verified result, though it marks it as self-reported by Anthropic (E3). The same sources say the ~72% human baseline was first crossed in February 2026, after the deadline (E4, E7). The best score reported by 30 Nov 2025 was therefore in the 65% to under-70% band, which the criteria define as mostly true.