METR: GPT-5.2 Task Horizon Estimate
Source: https://x.com (tweet) Author: METR (inferred) Date: 2025-2026
Summary
METR estimates GPT-5.2 with high reasoning effort has a 50%-time-horizon of approximately 6.6 hours on their expanded software task suite — the highest horizon estimate for any model to date. This is on their autonomy-focused evaluation.
Key Claims
- GPT-5.2 + high reasoning: 50% task horizon ~6.6 hrs (95% CI: 3hr 20min to 17hr 30min)
- Tested on METR’s expanded suite of software tasks
- This is the highest recorded horizon estimate
- METR’s methodology: measures how long a model can sustain autonomous useful work
Connection to Other Sources
Directly extends METR’s “task horizon doubling every 7 months” framework. GPT-5.2 at 6.6 hours represents a significant jump.
Concepts
- Autonomous Research — task horizon measures agent autonomy
- Coding Agents — software tasks as the primary domain