Enterprise AI in Numbers · Capability and cost
Length of tasks the best AI agents can complete
17.4 hours
17.4 hours (claude-mythos-preview-early) as of April 7, 2026, according to METR time horizons, via Epoch AI, up from 12.0 hours on February 5, 2026. The longest task, measured by how long it takes a skilled human, that the best measured model completes successfully half the time (METR's 50% time horizon), recorded each time a model sets a new record.
What it means for leaders
The length of task AI agents can handle on their own has been doubling in months, not years. Work that needed a person for a full day is coming into range, so redesign processes around supervised agents now rather than after the next jump.
What would change the reading: Records at higher success rates (80%), which matter more for production work.
The trend
Every reading
| Date | Value | Change | Detail | Source |
|---|---|---|---|---|
| 17.4 hours | +45.4% | claude-mythos-preview-early | Source | |
| 12.0 hours | +104.1% | claude-opus-4-6_unknown | Source | |
| 5.9 hours | +20.2% | gpt-5.2-2025-12-11_high | Source | |
| 4.9 hours | +30.6% | claude-opus-4-5-20251101 | Source | |
| 3.7 hours | +10.5% | gemini-3-pro-preview | Source | |
| 3.4 hours | +69.6% | gpt-5-2025-08-07_high | Source | |
| 120 minutes | +98.2% | o3-2025-04-16_unknown | Source | |
| 60 minutes | +55.7% | claude-3-7-sonnet-20250219 | Source | |
| 39 minutes | +89.3% | o1-2024-12-17_medium | Source | |
| 21 minutes | +1.0% | claude-3-5-sonnet-20241022 | Source | |
| 20 minutes | +78.1% | o1-preview-2024-09-12 | Source | |
| 11 minutes | +62.9% | claude-3-5-sonnet-20240620 | Source | |
| 7 minutes | +75.0% | gpt-4o-2024-05-13 | Source | |
| 4 minutes | — | gpt-4-0314 | Source |
Definition and method
- What it measures
- The longest task, measured by how long it takes a skilled human, that the best measured model completes successfully half the time (METR's 50% time horizon), recorded each time a model sets a new record.
- Population
- AI models measured by METR on its current task suite
- Updated
- When new models are measured
- Source
- METR time horizons, via Epoch AI · CC BY 4.0
Caveats
- The tasks are mostly software and research work; other domains may differ.
- A 50% success rate is far from the reliability most business processes need.
Cite this
Enterprise AI in Numbers, "Length of tasks the best AI agents can complete", https://lucianoayres.com/ai-brief/data/ai-agent-task-length/, retrieved October 25, 2026. Data: METR time horizons, via Epoch AI.