METR Time Horizons
metr.org- Category
- Developer Tools
- Rank
- No. 2569Tools index
Previous survey · No. 2428 ·
- Listed in
- Find AI benchmarks
- Type
- TOOL
- Use case
- Model & Agent Evaluation
- Date
About
A historical METR research dashboard estimating the human-expert task duration at which AI agents reach specified success probabilities, including 50% and 80%. It measures task difficulty on a defined suite, not AI wall-clock autonomy or every kind of work. The page states it is no longer actively updated, with a last update of 8 September 2026.
At a glance
METR Time Horizons is a research dashboard about the difficulty of software tasks AI agents can complete at specified success rates. It is worth reading to understand reliability tradeoffs and historical capability trends. It is no longer actively updated, with a stated last update of 8 September 2026, so it should not be your current model-shopping leaderboard.
Good fit
People interpreting claims that an agent can do hours of work. A horizon uses the estimated time a human expert would need for the task, not how long the AI runs unattended. Comparing 50% with 80% success makes that reliability condition visible instead of treating one duration as a promise about every task.
Consider something else when
You need a guarantee for your workflow, all recently released models or a measure of every kind of knowledge work. The task distribution is mainly self-contained software, machine-learning and cybersecurity work, with clear scoring criteria. Familiarity with a real project, messy requirements and human coordination can make a job quite different from this evaluation.
Start here
Read the update notice, keep the task-suite version fixed and switch the top chart from 50% Success to 80% Success. In the browser trial, its axis explanation and title changed to the 80% criterion. Check model dates and uncertainty before comparing points. The separate lower methodological panel has its own labels; do not assume every chart changed together.
Cost and requirements
The public graphs, methodology and linked data are readable without buying a product or supplying an API key. Reproducing an evaluation is a different undertaking involving model access, a tool scaffold, task infrastructure and repeated runs. Viewing the chart is not the same as testing an agent on your own application.
Limits to know
The page warns that measurements above 16 hours are unreliable with its task suite. The TH1.1 release explains that changing tasks and scaffolds can move estimates, with relatively few measured human baselines for long tasks. This review exercised a display control and read primary methodology; it did not reproduce the evaluation or establish future growth rates.
Sources checked 2026-10-30. Source-based assessment.
Why it made the leaderboard
Compare the task, benchmark version, harness, and grading method before using model scores to choose a model.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.