Skip to content
← Back to Agents & METR

METR's harder new tests put the best public AI at 40-hour tasks, up from 3 hours for GPT-5 in 2025

Task length is doubling about every 3.6 months, METR says, but its tests are getting too easy and too hard to score.

~40 hGPT-6 Astra's task length at 50% success (95% range 13–2,645 h)

Length of tasks AI agents finish half the time, METR's harder tests

METR 50% time horizon on its 222 older tasks plus 36 harder ones, at each model's release date. Log scale.

What this measures: How big a job an AI agent finishes on its own half the time. A task's length is how long a skilled human takes. This set adds harder tasks to METR's older ones, so it rates every model lower than the older chart.

New
1 min10 min1 h8 h40 h160 h20252026

Not comparable with the older METR chart: the harder tasks pull every model's figure down (Claude Opus 4.6: 6.4 hours here, 12 hours there).

METR's own fit to frontier models gives a doubling time of 3.6 months (95% CI 2.1–4.9). Each model's 95% range is very wide.

Source: METRData as of 21 Nov 2026
Show data table and sources
DateModelValue
13 May 2024GPT-4o7.1 min
5 Dec 2024o135 min
7 Aug 2025GPT-53 h
24 Nov 2025Claude Opus 4.54.6 h
5 Feb 2026Claude Opus 4.66.4 h
5 Mar 2026GPT-5.45.9 h
16 Apr 2026Claude Opus 4.78.4 h
23 Apr 2026GPT-5.511 h
28 May 2026Claude Opus 4.812 h
9 Jun 2026Claude Fable 516 h
9 Jul 2026GPT-5.6 Sol12 h
24 Jul 2026Claude Opus 520 h
1 Sept 2026Claude Fable 5.11.3 days
3 Sept 2026GPT-6 Astra1.7 days
22 Sept 2026Claude Opus 5.51.6 days

METR, an independent AI evaluation nonprofit, published its quarterly capability update on 20 November 2026, covering public models up to Claude Opus 5.5 and GPT-6 Astra. Recent models had saturated its Time Horizon 1.1 suite (222 tasks), so METR added TH2: 36 harder tasks in 12 families, such as optimisation contests and open-ended research. A model’s time horizon is the length of task, in skilled-human time, that it finishes half the time.

On the two suites pooled, GPT-6 Astra’s time horizon is about 40 hours and Claude Opus 5.5’s about 39. Claude Fable 5.1 is at about 30 hours and Claude Opus 4.8 at about 12. GPT-5 (August 2025) was at 3. Across models since GPT-4o (May 2024), METR finds the horizon doubling about every 3.6 months (95% CI 2.1–4.9), close to the 105 days it found on the older suite alone. On the verifiable half of TH2 (tasks scored automatically against a target), METR gives up to 27 hours for the best models and a faster doubling time of 1.65 months (95% CI 0.8–4.0).

METR paid specialist human baseliners $200–350 an hour, about $90,000 for 7 attempts, and found the models beat the baseliners on all 3 tasks it covered. On tasks needing judgment without clear feedback, agents were weaker than human experts. It says public evidence is not enough to rule out drastic AI R&D acceleration in the next couple of years. On a composite index of model ability (MCI), it sees no strong acceleration or slowdown across the public frontier.

What it means

METR’s old task-length chart ended in September because the best models outran it. On a harder set the growth continues at a similar pace, and the best public models do well at well-defined coding and optimisation work. Whether progress is speeding up is still open: METR sees hints in some subsets and not in the composite.

Caveats

  • The top figures are not separable: Astra and Opus 5.5 are within about 3% of each other, and each range spans 100× or more.
  • Only 9% of METR’s human-time estimates come from actual human baselines; the rest are its own estimates, and the results shift with them. Making TH2 times for high scores 4× longer lowers Astra to 32 hours.
  • Pooled figures are lower than on the older suite alone, so they cannot be set against earlier METR numbers: Claude Opus 4.6 is 6.4 hours here and was 12 hours on the older suite.
  • METR worries its easy-to-run tests resemble the environments labs train on, which could overstate real-world ability, and it found subtle reward hacking (gaming the scoring) by frontier agents.
  • The faster 1.65-month doubling on the verifiable subset may reflect an acceleration or other causes, such as a shorter date range, METR says.
Suggest a correction