Innovation & Tech

AI Task Horizons Double Every 110 Days, METR Finds

METR's November 2026 update: frontier model time horizons double every ~110 days and reach 27 hours on verifiable tasks, but research judgment and self-scoring lag the trend.

By Nathan Brooks

5 min read

METR Finds AI Time Horizons Doubling as Research Judgment Lags
METR Finds AI Time Horizons Doubling as Research Judgment LagsWalls.io / unsplash

What's News

  • METR reported on November 20, 2026 that frontier models' 50% time horizons double roughly every 110 days across TH1.1 and TH2.
  • On verifiable TH2 tasks, public frontier models reach time horizons of up to 27 hours.
  • Opus 5.5 as judge scored GPT-6 Astra's tasks about 0.2 points higher than the human scorer on a one-point scale.
  • Best autonomous Budget NanoGPT speedup was 15.7% in 48 hours versus a 41% human AI-assisted record.
  • Models placed at the 96th–98th percentile of human participants across 15 Al Zimmermann optimization contests.

Frontier AI models' time horizons are doubling roughly every 110 days, but their judgment on open-ended research work is not keeping pace, according to METR's November 20, 2026 evaluation update. The 50% time horizon — the human-estimated task duration at which a model succeeds half the time — kept its steep trend across METR's combined TH1.1 and TH2 test suites, covering public models through Opus 5.5 and GPT-6 Astra. METR typically evaluated 15 models on the combined suite.

The picture splits cleanly. Strong, measurable gains on verifiable tasks. Weaker performance on judgment, evaluation design and open-ended research — exactly the work that matters for automating AI R&D.

What does the harder suite show?

The new TH2 suite contains 36 tasks in 12 families. Eight are verifiable, with automatically scored objectives spanning optimization, machine learning, puzzles and games. Four cover conceptual reasoning, judgment and open-ended research with no cheap success check.

METR estimates a 95% confidence interval of 2.08 to 4.99-month ballpark for the combined doubling time — the reported range is 2.08 to 4.94 months — close to the previous 105-day trend for TH1.1 alone. Because TH2 is harder at any fixed human task duration, a model's 50% horizon on the combined suite sits lower than on TH1.1 alone.

On TH2's verifiable subset, the doubling time falls to 1.65 months, with a 95% confidence interval of 0.79 to 3.99 months. METR cautions this noisy result could reflect task composition, a restricted date range or selection during task development rather than broad acceleration. Less diverse task sets can also produce faster doubling, since longer tasks demand more of the same skills.

How were the humans calibrated?

Less directly than in METR's earlier work. Researchers defined score-and-time thresholds for 34 of the new tasks — three from successful human baselines, 31 from researcher estimates. Direct baselines informed only 9% of human-duration estimates. The specialists earned $200 to $350 per hour and could use Sonnet 5 for implementation. METR reports similar doubling times after adding noise to task-time estimates and substituting model-estimated completion times, though it remains unclear how comparable the baseliners were to frontier-lab researchers.

Strong scores, weaker judgment

On verifiable TH2 tasks, METR estimates current public frontier models' time horizons at up to 27 hours. On each of the three formally baselined tasks, the best model beat every recruited human baseliner.

Larger human pools tell a different story. Across 15 Al Zimmermann optimization contests, the strongest models placed at the 96th to 98th percentile of human participants in the median contest — roughly four humans still beat them. The conditions differ sharply: agents got 48 hours on a 28-core machine, while humans had about three months and their own hardware.

On Budget NanoGPT, the best autonomous-agent speedup was 15.7% in 48 hours. A recent AI-assisted human contributor delivered a 41% speedup over the previous record and estimated spending about 60 hours. The comparison is rough — human records build sequentially on earlier work — but it shows autonomous agents trailing the strongest human result while beating many other competent speedrunners.

The harder-to-verify tasks expose the real gap. In Making METR Tasks, agents had to design an informative new task for METR's own suite. METR classified some recent outputs as borderline useful, but none warranted inclusion. Models defaulted to automatically scored optimization tasks even when told to emphasize decisions with poor short-term feedback. Despite internet access, they did not investigate what METR would find useful, produce novel threat modeling, or build tasks on insightful hypotheses about model capabilities.

A related experiment tested whether models could score such outputs. Model judges agreed more with human ratings on weaker models' outputs, but gave inappropriately high scores to work from similarly strong or stronger models. Opus 5.5, the judge with the highest human agreement, scored GPT-6 Astra's tasks about 0.2 points higher than the human scorer — on a one-point scale.

METR says these results don't establish a universal judgment deficit. Models may be underelicited because they underused token limits and available affordances, and some showed good judgment on programmatically scored ML tasks.

What does this change for automation timelines?

METR is worried about measurement itself: containerized evaluations may resemble the reinforcement-learning environments used to train models, potentially overstating real-world capabilities. On cost, METR's Lifecycle Cost proxy — developing a task environment and scoring it 100 times — found the higher-cost subset 2.9 times more expensive than the lower-cost one, with models performing 2.9 times worse by time horizon, or 1.5 doublings. Improvement rates stayed similar on both subsets, so cost tracks difficulty without slowing progress.

The METR Capability Index shows no strong overall acceleration or deceleration across the public frontier, and developer-specific slope changes proved sensitive to methodological choices.

Still, METR says current public evidence can't rule out drastic AI R&D acceleration in the next couple of years — capabilities improving so fast that humans can no longer meaningfully understand, assess and react to catastrophic risks. Which capabilities would trigger that remains unknown, and easy-to-verify tasks alone might suffice. The upshot for practitioners: weaker judgment is a plausible constraint on research automation, not proof that acceleration must wait for models to master realistic, everyday work.

Source: unite.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,611 articles

Related articles

« Previous article