Read the evidence.
Put the work in context.
Source analysis Research
The duration uses a human baseline to describe task difficulty. Reliability, task selection and the dashboard’s update status determine what the chart can support.
AI-assisted editorial review · Sources checked · Recheck due
The takeaways
METR states that its time-horizon dashboard is no longer actively updated. It shows a last-update date of 8 September 2026. The measure estimates the human task duration at which an agent reaches a specified success probability. Its tasks primarily cover software engineering, machine learning and cybersecurity.
Source: METR: Task-Completion Time Horizons of Frontier AI Models
The 29 January TH1.1 release expanded the suite from 170 to 228 tasks and changed evaluation infrastructure. METR reported measured human baselines for only five of its 31 tasks lasting eight hours or more; the others used estimates. The release also discusses sensitivity to task composition.
Source: METR: Time Horizon 1.1
Editorial interpretation
Suppose a chart reports a two-hour horizon at 50% success. This is an illustrative number, not a current model score. Read it as a fitted prediction about tasks characterized by a two-hour human baseline within the evaluated distribution. It does not provide two hours of guaranteed unattended operation.
Now imagine using that headline to choose an agent for a weekly reporting job. Your job may involve finding a changed spreadsheet, resolving a missing value, obtaining a colleague’s decision and publishing a summary. A duration estimate alone does not tell you which of those steps the evaluation represented.
Write those steps down and identify which outcomes can be checked. For a first trial, you might evaluate the calculation and draft creation while leaving the colleague’s decision and publication with a person. That produces a bounded question you can actually investigate instead of treating an entire job as one benchmark task.
Editorial implications, not an independently measured result.
Editorial interpretation
Our proposed worksheet has five entries: the cited evaluation version, the task domain, the human baseline, the success threshold and the uncertainty around the estimate. Add the source’s last-update status so a useful historical measurement does not become a claim about every model available today.
Next, describe the consequence of an incorrect result in your own task. An occasional error in a private draft creates a different review burden from an error in a report sent to customers. Choose your acceptable outcome and verification method before deciding which model to trial.
For a superintelligence discussion, define the broader claim separately. A claim about general competence across unfamiliar domains requires evidence from those domains. A forecast about replacing a whole job also needs a description of the work that was omitted from the measurement. The chart can inform the discussion without supplying all of those missing premises.
Editorial implications, not an independently measured result.
The table scrolls horizontally on smaller screens.
| Entry | Question to write down |
|---|---|
| Scope | Which parts of my job resemble the evaluated tasks? |
| Consequence | What happens when this attempt fails? |
| Verification | Which output can I independently check? |
| Currency | When was the underlying source last maintained? |
Read alongside METR: Task-Completion Time Horizons of Frontier AI Models / METR: Time Horizon 1.1.
Keep the question open
Which part of your job matches the measured task domain, and what direct trial would resolve the remaining gap?
The evidence record
Source publication dates are reported as stated in the source record. An unstated publication date is not our check date.
Capabilities described by a vendor or paper do not establish performance in your environment. Our editorial standards explain the distinction.
Put the questions to work: plan your own agent evaluation ↗.