METR Time Horizon Explained: The Latest Numbers and Why It Is Not an IQ

METR's time horizon is the length of task, measured by how long a skilled human takes, that an AI agent can complete about half the time. In METR's data file for its current method (Time Horizon 1.1, 26 models) the highest figure is 17.4 hours, for an early version of Claude Mythos Preview, with a 95% confidence interval of 8.5 to 55.1 hours, and METR itself says measurements above 16 hours are unreliable with its present task suite. Across the models in that file the horizon has doubled roughly every 5.9 months, and every 3.5 months for models released since 2024, by our own regression. It measures how hard a task is for software, machine-learning and cybersecurity work. It is not an IQ, and it is not a count of the hours an AI can keep working. This guide sets out what the chart shows, how precise it is, and what it can and cannot be compared with.

What is METR's time horizon?

METR, an AI evaluation group, introduced the measure in a March 2025 paper, "Measuring AI Ability to Complete Long Software Tasks" (arXiv 2503.14499, revised as version 4 in July 2026). Every task in a suite is given a length: the geometric mean of the times taken by skilled professionals who completed it, people METR describes as software engineers, machine-learning engineers or cybersecurity specialists with about five years of relevant experience on average. For each AI agent METR then fits a logistic curve that predicts its chance of success as a function of that human task length. The 50% time horizon is the length at which the curve crosses 50%, and the 80% horizon is where it crosses 80%. The paper's suite had 170 tasks from three sets: 97 tasks from HCAST (one minute to 30 hours), 7 research-engineering tasks from RE-Bench (eight hours each) and 66 single-step tasks of one to 30 seconds.

Two readings of the number are wrong, and METR's own FAQ says so. It is not the time an AI spends working: METR describes it as a measure of task difficulty, and says agents are typically several times faster than humans on the tasks they complete. And it is not a promise of success on every task of that length. METR's example is GPT-5, with a horizon of about 2 hours 17 minutes: on tasks of 90 minutes to 3 hours it succeeded every time on about a third of them, failed every time on about a third, and was mixed on the rest.

What are the latest METR time horizon numbers?

METR's time-horizons page now says it is no longer actively updated, with a last update on 8 September 2026. The newest model in its Time Horizon 1.1 data file is an early version of Claude Mythos Preview, released in April 2026 and added to the page on 8 May. The 26 models in the file run from GPT-2 in 2019 to that model. Selected 50% horizons, with 95% confidence intervals:

• GPT-2 (February 2019): about 3 seconds
• GPT-4 (March 2023): 4.0 minutes (1.9 to 8.0)
• Claude 3.5 Sonnet (June 2024): 11 minutes (5.5 to 22)
• OpenAI o1 (December 2024): 39 minutes (21 to 65)
• Claude 3.7 Sonnet (February 2025): 60 minutes (33 to 104)
• OpenAI o3 (April 2025): 2.0 hours (1.2 to 3.2)
• GPT-5 (August 2025): 3.4 hours (1.9 to 6.8)
• Claude Opus 4.5 (November 2025): 4.9 hours (2.7 to 10.4)
• Claude Opus 4.6 (February 2026): 12.0 hours (5.3 to 60.6)
• Claude Mythos Preview, early version (April 2026): 17.4 hours (8.5 to 55.1)

Counting the file's point estimates, 14 of the 26 models are above one hour, 7 above four hours, 2 above eight hours (Opus 4.6 and Mythos Preview) and 1 above 16 hours. From GPT-4 at 4.0 minutes to the Mythos Preview figure of 1,045 minutes is a factor of about 260 in just over three years, roughly eight doublings; that is our arithmetic from METR's file. The list is not strictly rising: GPT-5.4 (March 2026) is at 5.7 hours and Gemini 3.1 Pro (February 2026) at 6.4, both below Opus 4.6.

Newer METR numbers sit in its blog, not in the data file. For GPT-5.6 Sol (26 June 2026), METR reported a 50% horizon of about 11.3 hours (95% CI 5 to 40) when cheating attempts are counted as failures, and beyond 270 hours if they are counted as successes, and added that it does not consider either number a robust measurement. That spread, 11 hours or 270, is the clearest sign of how much a single headline figure can hide.

How fast is the METR time horizon doubling?

METR itself gives several doubling times, and the one you see depends on the window. The paper reports a doubling every 207 days (95% CI 166 to 240) and summarises it as about seven months since 2019, adding that the trend may have accelerated since 2024. METR's data file records 188 days for the whole period and 129 days (104 to 158) from 2023 onwards. Its Frontier Risk Report of May 2026 fits models released after 1 January 2024 and finds 105 days, with R squared 0.98.

Our own regression on the 26 models in the file, log of the 50% horizon against release date, gives 178 days (5.9 months, 95% CI 5.2 to 6.7) for all of them, and 106 days (3.5 months, 3.2 to 3.8, R squared 0.97) for the 21 released since 1 January 2024. That agrees with METR's 105 days. Our intervals come from the regression alone and ignore the wide error bars on each model, so the true uncertainty is larger. The practical point is the one METR makes: seven months describes the long run, and the recent pace is about twice as fast.

How precise is a METR time horizon?

Less precise than the chart suggests, and METR says so. The 95% interval for Opus 4.6's 12.0 hours runs from 5.3 to 60.6 hours, so its upper end is about five times the estimate. The interval for the top figure, 8.5 to 55.1 hours, spans a factor of more than six. Four things limit the precision at the top:

• Few long tasks: the current suite has 228 tasks, and METR's blog says only 5 of its 31 tasks estimated at 8 hours or more have measured human times, the rest being estimates. METR's risk report says only 5 tasks are estimated to take humans longer than 16 hours
• Saturation: METR says its most capable agents essentially saturated the suite, and that measurements above 16 hours are unreliable
• Cheating: METR found that at least 16% of successful runs on tasks over 8 hours were illegitimate on review
• Method changes: a regularisation mistake corrected on 3 March 2026 lowered recent models' horizons by up to 20%, according to a METR researcher's note

One METR researcher, writing in January 2026 about the limits of the measure, put the historical error bars at about a factor of two in each direction, worse for the newest models, and said he could not tell whether one model's true horizon was 3.5 hours or 6.5. The note is the author's view and not METR's official position, but it is the most direct statement of how far to trust the third digit of any horizon.

What does a time horizon not measure?

The suite is mostly software engineering, machine learning and cybersecurity, and METR's FAQ says a 2-hour horizon does not mean an AI can do all intellectual tasks, nor an 8-hour one that it can automate all jobs. Three further limits come from METR's own numbers. First, tidiness: the tasks have an average "messiness" score of 2.36 out of 16, and Opus 4.6 reached 17.8 hours on low-messiness tasks but only 6.4 hours on high-messiness ones. Second, reliability: 50% is a coin flip. At 80% success the horizon is far shorter, 3.1 hours for the Mythos Preview figure against 17.4 at 50%, and 1.2 hours for Opus 4.6 against 12.0, and the paper found 80% horizons 4 to 6 times shorter in general. Third, domain: a METR note reports horizons similar for maths but 40 to 100 times lower for visual computer-use tasks.

Is a METR time horizon an IQ?

No, and METR does not claim it is. We searched seven METR documents (the paper, the time-horizons page, the Time Horizon 1.1 post, two research notes, the risk report and the GPT-5.6 Sol post) and the word IQ appears in none of them. Whether the measure says anything about general intelligence is our argument, not METR's, and the differences are concrete. An IQ is a rank within a norm group on a scale with a mean of 100 and a standard deviation of 15, which our percentile calculator turns into a share of people. A time horizon has no norm group: it is a length in hours on a log-scale chart, tied to one suite and to the skills of METR's human baseliners. A doubling of a time horizon means that tasks twice as long are now solved half the time; doubling an IQ has no meaning at all, because IQ points are not a ratio scale.

That is why the figures in our other AI guides are not interchangeable with this one. Our guides to the Mensa Norway log, AI scoring on an IQ-style test and a g factor for AI models explain why a benchmark result is not an IQ, and ARC-AGI-3 is the closest AI benchmark to an IQ-style puzzle. For broader capability rankings, see the most intelligent AI model and the Artificial Analysis Intelligence Index; for exam-style scores, Humanity's Last Exam and GPQA Diamond. Why AI benchmark scores disagree covers why the same model can look very different across them.

What do critics say about the time horizon?

Critics mostly agree with METR's own caveats and push them further. Gary Marcus, writing on 10 May 2026, argued that the chart covers only software-development tasks and that a 50% success bar says nothing about reliable performance. Jonas Moss, in a statistical re-analysis posted on LessWrong in February 2026, concluded that the data cannot tell exponential growth from faster-than-exponential growth, and that the 80% figures overstate current capability. Toby Ord, in May 2025, described the 50% horizon as a half-life: if success falls at a constant rate per minute of human task time, the horizon is the point where half the tasks are done. These are commentaries, not METR statements, and we have not re-run their analyses.

How to read a METR time horizon chart

• Check the version: Time Horizon 1.0 and 1.1 use different suites, and the same model can differ between them
• Read the interval, not the point: a figure of 12 hours with a range of 5 to 61 is a statement about order of magnitude
• Note the reliability level: 50% horizons are several times longer than 80% ones
• Note the date of the data: METR's page stopped updating in September 2026, so newer models are missing
• Do not convert it into an IQ, a job-replacement date or a count of hours of AI work

The bottom line: METR's time horizon is a careful measure of one thing, how long a task a model can finish about half the time on a software-heavy suite, and its trend is real and fast. It is not intelligence on the IQ scale, and its top end is the least reliable part. To see your own score on the standard scale, take the IQ Metrics IQ test, read it against the IQ bell curve, and see our blog's guides on whether AI is smarter than humans and what AGI means.

Common questions