×180
“×180. Since 2023, AI can complete tasks 180 times longer on its own.”
- Source
- METR, Time Horizon 1.1. Latest measurement added on May 8, 2026; since September 8, 2026, METR has stated that the page is no longer updated.
- Type
- Benchmark measurement, published by an independent research lab.
- Scope
- 50% task-completion horizon of frontier AI models at release, on software and research tasks. GPT-4 (March 2023): 3.99 minutes. o3 (April 2025): 119.7 minutes. Claude Opus 4.6 (February 2026): 718.8 minutes, or about 12 hours, the highest horizon measured for a publicly available model according to METR. A preliminary version of Claude Mythos Preview (April 2026) reaches 1,044.8 minutes (≈ 17 h 25 min; confidence interval: 8 h 29 min to 55 h 04 min).
- Method
- ×180 = 718.8 ÷ 3.99 (AI Origin calculation). Doubling time: 128.7 days since 2023 according to METR (confidence interval: 104 to 158 days), or about 4 months. For models released since 2024 alone, METR measures a doubling time of 105 days (Frontier Risk Report, May 2026).
- Caveat
- This is the length, in human expert time, of the tasks an AI completes successfully half the time. Not an AI that “works for 12 hours”. METR considers its measurements beyond 16 hours unreliable. No public model has reliably surpassed Opus 4.6: GPT-5.4 and Gemini 3.1 Pro, measured after it, remain below 6 h 30 min. The Mythos Preview data point exceeds that threshold, and the model is only available to Anthropic partners. Models released since then (Claude Opus 5.5, GPT-5.6 Sol…) have no reliable measurement: the task suite is practically saturated.
