The time-horizon metric was introduced by METR (Model Evaluation and Threat Research), a non-profit that evaluates frontier AI systems, in the March 2025 paper “Measuring AI Ability to Complete Long Software Tasks”, with Thomas Kwa and Ben West listed first among 26 authors. Rather than another score out of 100, it expresses AI progress in a unit anyone can understand: hours of human work.
What it measures
The 50% time horizon is the length of task, measured by how long it takes skilled human professionals, that an AI agent can complete with a 50% success rate. If a model’s time horizon is two hours, it succeeds about half the time on tasks that take an expert around two hours, more often on shorter tasks and less often on longer ones. METR also reports an 80% horizon, which is much shorter, for situations where reliability matters more.
How it works
METR assembled a suite of software engineering, machine-learning research and cybersecurity-style tasks of very different lengths, from seconds to many hours. It combines task collections including RE-Bench, HCAST and a set of 66 novel shorter tasks. Skilled humans attempted the tasks to establish how long each one takes. AI agents then attempted the same tasks.
For each model, METR fits a curve of success probability against human task length and reads off the length at which predicted success drops to 50%. Results come with wide confidence intervals. The current version of the suite, Time Horizon 1.1, follows the same method as the original paper with a larger set of tasks.
A typical long task might be to optimise a machine-learning training script to hit a target speed, or to implement and debug a small software system from a specification, work that would take a human engineer several hours.
Why it matters
The paper found that the time horizon of frontier models had been doubling roughly every seven months since 2019, and that the trend may have accelerated in 2024; at the time, models like Claude 3.7 Sonnet had a horizon of around 50 minutes. METR’s current data file estimates a doubling time of about 129 days (roughly four months) for models from 2023 onward. Because the metric is intuitive and extrapolates naturally, it has become one of the most cited charts in debates about how quickly AI agents might automate real work.
Limitations and criticisms
- Mostly software tasks. The suite focuses on software, ML and related technical work. Time horizons on other kinds of tasks may be very different.
- 50% is not reliable. Succeeding half the time is far from what is needed for unsupervised use; 80% horizons are much shorter.
- Wide uncertainty. Confidence intervals often span a factor of two or more. METR states that measurements above 16 hours are unreliable with its current task suite.
- Extrapolation risk. Projecting the doubling trend years ahead assumes it continues; real-world tasks are messier than benchmark tasks.
- No longer actively updated. METR’s time-horizons page, last updated on 8 September 2026, says it is no longer actively updated.
Reading the scores
The values above are METR’s p50 estimates on Time Horizon 1.1, converted from minutes to hours, as published on its page (last updated 8 September 2026). The top estimate exceeds 16 hours, which METR itself flags as unreliable, and the confidence intervals are wide: for Claude Mythos Preview (early), roughly 8.5 to 55 hours. For comparison, METR estimated GPT-5 at around 2 hours 17 minutes. Treat these as rough orders of magnitude, not precise rankings.