METR 50% Task-Completion Horizon
A fitted 50 percent success threshold for METR benchmark tasks, not elapsed agent runtime
Where the METR 50% Task-Completion Horizon stands now
The METR 50% Task-Completion Horizon was 17.4 hours in April 2026, the highest in data back to 2023. It has been higher than a year earlier for five releases in a row.
How the METR 50% Task-Completion Horizon has moved since 2019
METR 50% Task-Completion Horizon: up since 2019, the highest in data back to 2023
Hours of human-expert task time at 50% predicted success, by model release date; METR says values above 16 hours are unreliable: 17.4 hours in April 2026, up from 2 hours a year earlier. Event-driven, 2019–2026.
What the METR 50% Task-Completion Horizon measures
Measurement basis: METR's 50 percent time horizon is the estimated human-expert task duration at which the evaluated agent's fitted success probability is 50 percent; it is not elapsed agent runtime. The current task suite primarily covers self-contained, well-specified software engineering, machine learning, and cybersecurity tasks with clear success criteria, and METR says measurements above 16 hours are unreliable.
METR 50% Task-Completion Horizon by release, 2019 to 2026
| Release | Hours | Change from a year earlier |
|---|---|---|
| Apr 2026 | 17.4 hours | — |
| Feb 2026 | 12 hours | — |
| Dec 2025 | 5.9 hours | — |
| Nov 2025 | 4.9 hours | — |
| Aug 2025 | 3.4 hours | — |
| May 2025 | 1.7 hours | — |
| Apr 2025 | 2 hours | — |
| Feb 2025 | 1 hour | — |
| Dec 2024 | 0.7 hours | — |
| Oct 2024 | 0.3 hours | — |
| Sep 2024 | 0.3 hours | — |
| Jun 2024 | 0.2 hours | — |
| May 2024 | 0.1 hours | — |
| Apr 2024 | 0.1 hours | — |
| Mar 2024 | 0.1 hours | — |
| Nov 2023 | 0.1 hours | — |
| Mar 2023 | 0.1 hours | — |
| Mar 2022 | 0 hours | — |
| May 2020 | 0 hours | — |
| Feb 2019 | 0 hours | — |
Event-driven · METR (Model Evaluation & Threat Research)
Counties with the highest labor scores
There's no county figure for this measure, so this list uses the labor part of our County Distress Index.
| Rank | County | Score (0–100) |
|---|---|---|
| 1 | Apache County, AZ | 99.2 |
| 2 | Mingo County, WV | 98.7 |
| 3 | Navajo County, AZ | 97.3 |
| 4 | Luna County, NM | 97.2 |
| 5 | McKinley County, NM | 96.8 |
| 6 | Lake County, CA | 96.8 |
| 7 | Clare County, MI | 96.6 |
| 8 | Harlan County, KY | 96.2 |
| 9 | Wyoming County, WV | 96.2 |
| 10 | Pike County, KY | 96.1 |
Questions about how long AI can work on its own
What is METR's time horizon?
It is the length of a task, measured by how long human experts take, at which METR predicts an AI agent will succeed half the time. The latest reading on this page is 17.4 hours, for a model released in April 2026. It measures task difficulty, not how long the AI works.
Does this mean AI can do a full day of someone's job?
No. METR's tasks are mainly self-contained software engineering, machine learning and cybersecurity problems with automatic scoring, done by experts with little context. METR says its human times likely overstate how long an expert takes on the job, and its separate study found much shorter horizons for visual computer-use tasks.
How reliable is the latest number?
The latest reading on this page is 17.4 hours. METR says values above 16 hours (960 minutes) are unreliable on its current tasks. METR's own error bars have been about a factor of two each way, and it revises past estimates when it changes its method.
Where does this data come from?
METR, a research nonprofit, publishes time horizons for the AI models it chooses to evaluate. It says its list is not a complete record of the most capable models, and each point on the chart is dated by the model's release, not by when METR measured it.
Sources and method for the METR 50% Task-Completion Horizon
- METR (Model Evaluation & Threat Research) Measurement basis: METR's 50 percent time horizon is the estimated human-expert task duration at which the evaluated agent's fitted success probability is 50 percent; it is not elapsed agent runtime. The current task suite primarily covers self-contained, well-specified software engineering, machine learning, and cybersecurity tasks with clear success criteria, and METR says measurements above 16 hours are unreliable. We use the event-driven figures as published, from February 2019 through April 2026, with no adjustment of our own. Model Evaluation & Threat Research (METR)’s report
We write with assistance from AI tools. Figures come from the data and sources named on this page.