Key takeaways

  • METR's 50% time horizon is the human task length at which an AI agent succeeds half the time, found by fitting a logistic curve to success against human task time.
  • METR's November 20, 2026 update puts GPT-6 Astra at 40 hours and Claude Opus 5.5 at 39 hours, with very wide confidence intervals, and a doubling time of about 3.6 months.
  • New judgment-heavy tasks show models still trail experts: no agent-built evaluation task was useful enough for METR to include.

The METR time horizon is the length of a task, measured in how long a skilled human takes to do it, that an AI agent can finish with a 50% success rate. In its November 20, 2026 update, METR put Claude Opus 5.5 and GPT-6 Astra at about 39 and 40 hours, and said the number keeps doubling roughly every 3.6 months.

What METR is

METR, short for Model Evaluation and Threat Research, is a nonprofit that tests frontier AI models. Its stated mission is to "discover and publish evidence about the capabilities of advanced AI systems." METR says it has not accepted funding from AI companies, though it does receive free tokens from them for its evaluations.

Its November 5 report estimating that Anthropic's progress had roughly doubled in speed is a separate piece of work by a different team, covered here earlier.

How the METR time horizon is computed

The method comes from the March 2025 paper Measuring AI Ability to Complete Long Tasks (Kwa, West, Becker and 23 co-authors). It works in three steps.

  1. Time the tasks with humans. METR built a suite of 170 tasks, mostly software engineering, machine learning and cybersecurity: 97 tasks from HCAST, 7 from RE-Bench and 66 very short ones. Paid experts did the tasks, and each task's length is the geometric mean (a multiplicative average) of the successful humans' times. Tasks ran from a few seconds to about 30 hours.
  2. Run the AI agent on every task. Each attempt is scored as a success or a failure.
  3. Fit a curve. For each model, METR fits a logistic curve, an S-shaped line, of success rate against the log of human task time. Models succeed on short tasks and fail on long ones. The task length where the curve crosses 50% is the model's 50% time horizon. The same curve also gives an 80% horizon, which the paper found was roughly five times shorter.

In the paper, Claude 3.7 Sonnet had a 50% horizon of about 50 minutes. Across models since 2019, horizons doubled about every 207 days, or roughly seven months, with signs of a faster pace in 2024 and 2025.

What the November 20 update found

The old suite, now called Time Horizon 1.1, ran out of headroom. METR says that within a year of the March 2025 launch, models had saturated it, and its own time horizons page had already warned that measurements above 16 hours were unreliable. So METR added a new suite, TH2: 36 tasks in 12 families, including optimization contests, a NanoGPT training speedrun, games like Slay the Spire, and four families of judgment-heavy work.

The main numbers:

Measure Result
GPT-6 Astra, 50% horizon (TH1.1 + TH2) 40 hours (95% CI 13 to 2,645)
Claude Opus 5.5, 50% horizon (TH1.1 + TH2) 39 hours (95% CI 12 to 1,253)
Best horizon on the verifiable part of TH2 Up to 27 hours
Doubling time, TH1.1 + TH2 3.61 months, or 110 days (95% CI 2.08 to 4.94 months)
Doubling time, TH2 verifiable tasks only 1.65 months (95% CI 0.79 to 3.99)
Previous TH1.1 doubling time 105 days

METR also mentions Fable 5.1 at about 30 hours and Opus 4.8 at about 12 hours. On a naive extrapolation, METR says models would reach a 50% horizon of 200 hours, the longest task in its set, in July 2027.

The 1.65-month figure on the new verifiable tasks looks like a speed-up, but METR doesn't read it that way. It notes that the MirrorCode suite showed similarly fast doubling in late 2025, and that TH2 has few task families and a short date range. A separate check with a statistical model it calls the METR Capability Index showed no obvious acceleration across all public frontier models.

Where models still fall short

The judgment tasks are the newest part. As MIT Technology Review put it, models lag on judgment compared with coding.

  • Designing evaluations. Agents were asked to build a task METR would find informative. None of the results were useful enough to include. Agents mostly built auto-scored optimization tasks, even when told to make tasks where early feedback is a poor guide to success.
  • Judging other models. Model judges rated tasks made by stronger models too generously. Opus 5.5, the best judge, scored GPT-6 Astra's tasks about 0.2 points (out of 1) higher than METR's human rater.
  • Rating philosophy critiques. Many models beat experts who spent 5 minutes on the task, but every public model trailed experts who spent 15 minutes.
  • Top humans. In 15 long optimization contests, the strongest models placed at the 96th to 98th percentile, but in the median contest about four humans still beat them.

METR also found models do worse on tasks that are expensive to build and score, even at the same length. Its higher-cost subset was 2.9 times as expensive, and model horizons on it were 2.9 times shorter. Progress on both subsets ran at a similar rate.

What changed in the method

Three things. First, human timing is mostly estimated now. Only 9% of TH2 time estimates come directly from human baselines; the rest come from public records, experts and extrapolation, because few people can do a 100-hour task on demand. Baseliners were also allowed to use Sonnet 5 for implementation.

Second, scoring is richer. Instead of one pass-or-fail threshold, many TH2 tasks have several score levels, each with its own estimated human time. Some are graded by humans or by model judges.

Third, METR deliberately added tasks it expects to be unlike the training environments labs use, because it worries reinforcement learning on similar tasks inflates scores.

Limits and common misreadings

The time horizon is a narrow measure, and METR's own FAQ lists the mistakes people make with it.

  • It is not how long the model works. A 40-hour horizon means the task would take a human 40 hours. The agent may finish much faster, or work for a while and fail.
  • It is a coin flip, not reliability. 50% success is the definition. The 80% horizon is much shorter, and METR does not report 99%.
  • It is not job automation. The tasks are cleaner and better specified than real work. METR's own explainer says an 8-hour horizon does not mean AI can do 8-hour jobs.
  • It is mostly software. The suites are built from coding, ML and security tasks. Other domains give different numbers.
  • The error bars are huge. GPT-6 Astra's 40 hours has a 95% interval from 13 to 2,645 hours. The point estimate is rough.

A fair reading: on well-scoped, checkable technical work, frontier models now handle tasks that take a skilled person about a work week, half the time. On work that needs judgment about what is worth doing, models still trail experts. Technology Review reports that Opus 5.5 now lands "borderline above the bar" on evaluation design, where a year ago every model was clearly below it.

METR's open questions are which capabilities AI research actually needs, and whether hard-to-verify skills are improving more slowly. The full update, with appendices, is on METR's blog.

Frequently asked questions

What is METR?

METR (Model Evaluation and Threat Research) is a nonprofit that evaluates frontier AI models and publishes evidence about their capabilities. It says it has not accepted funding from AI companies, though it receives free tokens from them for evaluations.

Does a 40-hour time horizon mean an AI can work for 40 hours on its own?

No. It means the model succeeds about half the time on tasks that take a skilled human about 40 hours. The agent may finish faster, and it fails the other half of the time.

How fast is the METR time horizon growing?

The original 2025 paper found a doubling time of about seven months since 2019. METR's November 2026 update measures about 3.6 months (110 days) across its old and new task suites.

What did the METR report find about AI judgment?

On new hard-to-verify tasks, models trailed experts. None of the evaluation tasks agents built was useful enough for METR to include, and every public model trailed experts who spent 15 minutes rating philosophy critiques.

Sources

  1. AI R&D-relevant model capabilities are improving rapidly and are getting harder to measureMETR
  2. Measuring AI Ability to Complete Long TasksarXiv
  3. Task-Completion Time Horizons of Frontier AI ModelsMETR
  4. About METRMETR
  5. The most misunderstood graph in AI just got a major updateMIT Technology Review