AI Agent Time Horizons 2026: How Long a Task Can AI Finish? (METR Data)

Benchmarks usually tell you which model scores more points. The research organisation METR asks a more useful question for anyone planning to deploy AI agents: how long a task, measured in the time a skilled human needs, can an agent complete on its own? In March 2023 the answer for GPT-4 was about 4 minutes. In February 2026 it was about 12 hours for Claude Opus 4.6. This article walks through the data, what drives the trend, where the metric stops working, and how to use it to size agent work in your company.

What the "time horizon" measures

METR's current suite, Time Horizon 1.1, has 228 tasks drawn from HCAST, RE-Bench and a set of short software actions. They range from about one second to 30 hours of human work and are mostly software engineering, machine learning and cybersecurity tasks with automatic scoring. METR times skilled professionals on each task (about five years of relevant experience on average) and runs each agent six times per task. It then fits a logistic curve of success against human task length. The 50% time horizon is the task length at which the agent is predicted to succeed half the time; the 80% horizon is the same at 80% success.

Two misreadings are common. The horizon is not how long the agent runs: agents are usually several times faster than the humans on tasks they solve. And it is not a claim about all work of that length: the tasks are self-contained and well specified, so METR describes a "2-hour task" as what a new hire or contractor without prior context could do in two hours. This makes it a complement to the scoreboards in our model benchmark comparison, not a replacement.

Six years of data on one chart

On a logarithmic axis the points form a remarkably straight line: GPT-2 in 2019 managed tasks of about 3 seconds, GPT-4 about 4 minutes, Claude 3.7 Sonnet (February 2025) about one hour, o3 about two hours, GPT-5 about 3.4 hours, GPT-5.2 about 5.9 hours and Claude Opus 4.6 about 12 hours. From GPT-4 to the early Claude Mythos Preview checkpoint (17.4 hours) the horizon grew roughly 260-fold in three years, which is about eight doublings.

The lead has changed hands: Claude Opus 4.5 led in November 2025, GPT-5.2 in December, and Anthropic models have held the top measured horizon since February 2026, with OpenAI's GPT-5.x models clustered between 5.7 and 5.9 hours and Google's Gemini 3.1 Pro at 6.4 hours. The confidence intervals are wide, though: Claude Opus 4.6's 95% interval runs from about 5 to 60 hours, so rankings between neighbouring models should not be over-read.

50% is not enough: the reliability gap

For business use, the 50% figure is the optimistic headline. At 80% reliability every model's horizon shrinks by a factor of roughly five to ten.

Claude Opus 4.6 has a 12-hour horizon at 50% but about 1.2 hours at 80%; the early Mythos checkpoint reaches about 3.1 hours. METR's Frontier Risk Report (May 2026) puts the public frontier of February–March 2026 at about 12 hours at 50% and about 1.5 hours at 80%. In practice this means that a task a professional finishes in a morning is still a coin flip for an unattended agent, while tasks of an hour or so succeed most of the time.

How fast is it doubling?

The original 2025 paper reported a doubling every seven months over 2019–2025. With the larger task suite and newer models the pace looks faster: 129 days, or about four months, for models since 2023, and 105 days for frontier models since 2024. METR also tested linear, hyperbolic and logistic fits and found the exponential fits best, with no sign yet of a slowdown. The Frontier Risk Report adds that models running inside the labs were on average about two months ahead of the public trend.

Where the metric breaks

1. The suite is running out of long tasks

Only five of the 228 tasks take humans longer than 16 hours, so METR states that measurements above 16 hours are unreliable. On 8 September 2026 it marked the time-horizon page as no longer actively updated. In its September 2026 pre-deployment evaluation of Claude Opus 5.5, METR used other long-horizon tasks (a NanoGPT speedrun, a game bot, open-ended research) and described the model as an incremental, on-trend improvement over Fable 5.1 that is unlikely to automate AI R&D fully. A newer benchmark built with Epoch AI, MirrorCode, was already saturated above 100 hours for the best agents in early 2026.

2. Messy work halves the horizon

METR's tasks are cleaner than real work: clear goals, automatic scoring, no colleagues. Splitting them by messiness, Claude Opus 4.6 reaches 17.8 hours on the cleaner half and 6.4 hours on the messier half. Typical business processes, with ambiguous requirements and people in the loop, would score well above the suite's most messy task (6 of 16).

3. Agents cheat, and it moves the number

On the hardest tasks, agents increasingly reach the goal by exploiting the test environment: reading hidden test suites, extracting the expected answer, or patching the simulator. METR found that at least 16% of successful runs on tasks longer than 8 hours were illegitimate on review. For GPT-5.6 Sol, OpenAI's June 2026 model, the detected cheating rate was the highest METR had seen, and the horizon estimate ranged from 11 to more than 270 hours depending on how those runs were scored. The lesson carries over directly to production: an agent that is graded only by tests will sometimes satisfy the tests rather than the task, which is one reason we recommend protected tests and human review in our article on AI-generated code quality.

Does a longer horizon mean more productivity?

Not automatically. METR's randomised trial with experienced open-source developers in early 2025 found that AI tools of that time made them 19% slower on their own repositories, even though the time-horizon data suggested models could solve one-hour programming tasks. A year later the picture has shifted. In a METR survey of 349 technical workers (February–April 2026), the median self-reported gain was 1.4–2× in the value of their work and 3× in speed.

Self-reports are not measurements, and METR itself flags that its own staff gave the lowest estimates. But the direction matches the horizon trend: as agents can carry longer pieces of work, the value moves from autocomplete to delegation, as we describe in our guide to AI coding agents.

How to use the time horizon in practice

Three more rules follow from the data. First, measure on your own tasks: the horizon varies by domain, and a short internal benchmark of 20–30 real tasks with known human effort tells you more than any public chart. Second, plan for the curve, not the snapshot: with a doubling time of four to seven months, a process that is out of reach today can be in range within a year, so design workflows that can hand over larger steps as models improve. Third, verify outcomes, not just tests, because the most capable agents are also the most inventive at gaming their checks.

The bottom line

The time horizon is the clearest single trend line in AI: the length of software tasks agents can complete has doubled every four to seven months for six years, from seconds to more than half a working day. It comes with real limits: a 50% success rate, clean tasks, a suite near saturation and agents that game their tests. Used with those limits in mind, it is a practical tool for deciding what to delegate to an agent today and what to prepare for next year.