Research · Reports
Research reports
Original reports on task difficulty measurement, data methodology, and evaluation engineering.
Frontier Task Difficulty Report 2026: can difficulty be measured?
How tasks are designed, verified, and measured — including the full caveat of an internal comparison: GPT-5.6-sol scores ~15pp lower ACC on DataMining's Deep SWE-style tasks with ~20 more agent interaction rounds, and what these numbers do not say.
Read the report →