METR has rebuilt its time horizon plot by adding judgment-heavy tasks, revealing that while AI capability growth still appears exponential, models remain poor at designing evaluations themselves. The update expands the suite to 228 tasks and adjusts estimates, suggesting a doubling rate closer to three to four months post-2024.

METR has rebuilt its time horizon plot, the exponential chart the AI industry watches like a markets feed, and the update carries a double message: capability growth still looks exponential, and the models remain poor at one task in particular - designing the evaluations themselves.
Grace Huckins reports at MIT Technology Review that the nonprofit has added judgment-heavy tasks to the suite behind the plot. The finding that drew the most attention is that models lag at designing AI evaluations, a task METR now measures the way it measures coding or cybersecurity work.
The chart, for the fourth time
For readers who have been avoiding it: METR (Model Evaluation & Threat Research) does not score models on accuracy. It scores them on task length. Every task in its suite is labeled with the time a skilled human takes to complete it. For each model, METR fits a logistic curve of success probability against that human-time axis, and reads off the point where the model succeeds half the time. That is the 50% time horizon. In March 2025 the original paper put the doubling rate near seven months. In January, the Time Horizon 1.1 update expanded the suite to 228 tasks and moved several estimates; observers noted the post-2024 trend looked steeper, closer to a doubling every three to four months.
The plot became the industry's informal clock for AGI timelines. Sequoia used it in a post titled "2026: This is AGI." Investors cite it. Governments quote it. METR staff have spent the intervening months explaining what it does not say.
What judgment-heavy tasks change
The original suite was software-heavy by construction: HCAST engineering tasks, RE-Bench research-engineering environments, short computer-use items. METR's own FAQ concedes the tasks are "self-contained and well-specified, with clear success criteria that can be automatically evaluated" - the opposite of most real work, where the objective is fuzzy and success requires judgment rather than a passing test.
Adding judgment-heavy tasks pushes the metric toward that messier distribution. METR's follow-up work has long shown that horizons differ sharply across domains; capabilities are jagged relative to humans. A plot that averages exponential growth across software and judgment is a different claim than the old software-only plot - and a smaller claim than the AGI readings it inspired.
The awkward result: models cannot write the tests
The models-lag-at-designing-evaluations finding is the one with consequence beyond the chart. The field is increasingly outsourcing eval design to the models themselves - generating tasks, judging outputs, scaling human-preference pipelines. Recursive improvement of AI systems depends on AI systems producing reliable tests. If the same models that clear four-hour engineering tasks cannot design good evaluations, then every layer of automated eval infrastructure sits on a known weak capability.
METR has been here before. Its earlier research flagged that models score worse on "messy" tasks, and that its estimates above roughly 16 hours were statistically unreliable on its own task suite. Each fix has tightened the measurement and shrunk the overclaiming.
The honest summary of the update: some capabilities keep growing exponentially on a better-spread task suite, the chart's meaning keeps getting narrower, and the least impressive result is precisely the one the industry was hoping to automate away.
Sources
- Techmeme summary: METR updates its time horizon plot to add judgment-heavy tasks (Grace Huckins / MIT Technology Review)
- MIT Technology Review: This is the most misunderstood graph in AI
- MIT Technology Review: METR's time horizon plot gets a major update
- METR: Time Horizon 1.1
- METR: Task-Completion Time Horizons of Frontier AI Models
Frequently asked questions
What does METR measure on its time horizon plot?
METR measures task length based on the time a skilled human takes to complete a task, fitting a logistic curve of success probability against that human-time axis to determine when a model succeeds half the time.
How have the new judgment-heavy tasks changed METR's findings?
The original suite was software-heavy and well-specified; adding judgment-heavy tasks reveals that models lag significantly at designing AI evaluations, a critical weakness for recursive system improvement.
What is the current estimated doubling rate for AI capabilities?
In January 2025, observers noted the post-2024 trend looked steeper, closer to a doubling every three to four months, up from the seven-month rate cited in March 2025.
