Why this story
- Why it matters
- The comparison gives researchers a way to track increasingly long tasks while exposing how little a benchmark says about choosing the right experiment when feedback is poor.
- Attention gap
- Ori’s seven-day headline-index search recorded 0 of 11 outlets with at least one relevant headline. The publisher located METR’s original report. Targeted panel searches did not establish additional coverage. Index coverage varies by outlet.
In a November 20 report, the nonprofit evaluator METR estimated how much human working time corresponds to a task a frontier AI model can complete with a 50% chance of success. Across two combined test suites, that horizon doubled about every 3.6 months. Its 95% confidence interval was roughly 2.1 to 4.9 months.
The tasks included coding and optimization. The models often matched or beat the skilled people METR recruited. Human experts generally showed better judgment on tasks with weak or delayed feedback. Only 9% of the study’s estimates of human completion time came directly from baselines measured with people.
METR said the evidence leaves open the possibility of rapid acceleration in AI research over the next few years. It also reported substantial uncertainty about which skills such acceleration would require. The benchmark does not establish a timetable for automating research jobs.
What the evidence cannot tell us
- The findings come from METR’s selected benchmark tasks and its own report. Its extrapolations depend on the trend continuing.
What to watch
Check whether the trend persists in new models and on research tasks with less immediate feedback.
Read the sources. Credit the work.
The reporting and documents behind this brief, linked directly to the source. Verified publisher and author credits appear where available; other entries show their actual website address. Several links may draw on the same underlying reporting or research.
METR
AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure (opens in a new tab)metr.org · November 20, 2026
Primary document / sourceContent read for this publication · November 22, 2026Publisher credit checked November 22, 2026. View credit on the source page (opens in a new tab).
This story was written and checked using AI against the linked evidence. No routine human review is claimed. Read the method.
Suggest an update or correctionPublished corrections