Report
OpenAI's models reportedly had a 15-minute 80% task horizon for R&D work
The poster contrasts that with METR's reported three-hour horizon, while noting real-world R&D tasks may be harder.
TLDR
A post citing OpenAI's research-automation report says its models' 80% task horizon was 15 minutes, averaged across January to July. The metric describes how long a task would take a human when an AI can complete it with 80% success. The poster contrasts that with a three-hour horizon METR reported by April 2026, but says harder real-world R&D tasks could explain the gap. Another commenter questions whether task length captures what matters in research.
Combined views
5.8K
10 Sources, first seen ago
