AI Models & Platforms
METR Finds AI Time Horizons Doubling as Research Judgment Lags

Key PointsFrontier models' time horizons double roughly every 110 days but research judgment lags…
View more
- Frontier models' time horizons double roughly every 110 days but research judgment lags behind verifiable gains.
- Models placed at the 96th to 98th percentile in optimization contests yet trailed top humans on autonomous speedups.
- Model judges overrated strong models' research tasks by about 0.2 points on a one-point scale.
METR reported on November 20, 2026, that public frontier models’ 50% time horizons—the human-estimated task duration at which a model is predicted to succeed half the time, not how long the model itself works—are doubling roughly every 110 days across its combined TH1.1 and TH2 evaluations. The expanded testing shows strong gains on verifiable work but weaker judgment on open-ended research and evaluation design.
The update covers public models through Opus 5.5 and GPT-6 Astra and examines capabilities relevant to automating AI research and development (R&D). METR typically evaluated 15 models on TH1.1 and TH2. The new TH2 suite contains 36 tasks in 12 families: eight verifiable families with well-defined, automatically scored objectives across optimization, machine learning, puzzles and games, and four families covering conceptual reasoning, judgment and open-ended research without a cheap check for success.
A Harder Suite with a Similar Growth Trend
METR estimates a 95% confidence interval of 2.08 to 4.94 months for the combined doubling time, close to the previous 105-day trend for TH1.1 alone. Because TH2 is harder at a fixed human task duration, a model’s 50% time horizon on the combined suite is lower than on TH1.1 alone.
On TH2’s verifiable subset, the estimated doubling time falls to 1.65 months, with a 95% confidence interval of 0.79 to 3.99 months. METR cautions that this noisy result could reflect task composition, a restricted date range or selection during task development rather than broad acceleration. Less diverse task sets may also produce faster doubling because longer tasks require more of the same skills.
Human calibration is less direct than in METR’s earlier work. Researchers defined score-and-time thresholds for 34 of the new tasks, using successful human baselines for three and researcher estimates for 31; direct baselines informed only 9% of human-duration estimates. The specialists were paid $200 to $350 per hour and could use Sonnet 5 for implementation. METR reports similar doubling-time results after adding noise to task-time estimates and substituting model-estimated human completion times, though it remains uncertain how comparable the baseliners were to frontier-lab researchers.
Strong Scores, Weaker Research 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 outperformed every recruited human baseliner.
Larger pools of human data show that top experts can still finish ahead. Across 15 Al Zimmermann optimization contests, METR reports that the strongest models placed at the 96th to 98th percentile of human participants in the median contest, while roughly four humans beat them. Agents had 48 hours on a 28-core machine; human participants had about three months and their own hardware, so the rankings depend on those different conditions.
On Budget NanoGPT, METR reports a best autonomous-agent speedup of 15.7% in 48 hours, compared with a 41% speedup over the previous record by a recent AI-assisted human contributor who estimated spending about 60 hours. The comparison is rough because human records build sequentially on earlier work, but it shows autonomous agents trailing the strongest human result while exceeding many other competent speedrunners.
The hard-to-verify tasks reveal a different limitation. 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 in the suite used in the report. Models generally created automatically scored optimization tasks even when instructed to emphasize decisions with poor short-term feedback. Despite internet access, they did not independently investigate what METR would find useful, produce novel threat modeling, or base tasks on particularly insightful hypotheses about model capabilities. Those ratings rest on METR’s own research and threat-modeling judgments.
A related experiment examined whether models could score those outputs reliably. METR found that model judges agreed more closely with human ratings when assessing weaker models’ outputs, but awarded 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.
These results do not establish a universal judgment deficit. METR says models may be underelicited on judgment tasks because they underused token limits and available affordances, and notes that frontier models displayed good judgment of unclear generality on some programmatically scored machine-learning tasks.
Costlier Evaluations and the Acceleration Question
The measurement problem extends beyond judgment. METR is concerned that containerized evaluations may resemble the reinforcement-learning environments used to train models, potentially overstating real-world capabilities as similar environments become common in training.
Lifecycle Cost is a rough, model-estimated proxy for developing a containerized task environment and running and scoring it 100 times, rather than a direct estimate of an AI developer’s training-environment costs. After controlling for task length, METR found the higher-cost subset was 2.9 times as expensive as the lower-cost subset on average, while models performed 2.9 times worse by time horizon, equivalent to 1.5 doublings. Models nevertheless improved at similar rates on the two subsets, so the data associates cost with difficulty without showing slower progress on costlier tasks.
METR reports no strong overall acceleration or deceleration across the tested public frontier on its separate METR Capability Index, which summarizes model performance on a common scale. Developer-specific slope changes were sensitive to methodological choices, leaving their validity uncertain.
METR nevertheless says current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years, defined as capabilities improving so quickly that humans can no longer meaningfully understand, assess and react to catastrophic risks from new systems. Which capabilities would produce that outcome remains unknown, and easy-to-verify tasks alone might suffice. The findings therefore make weaker judgment a plausible constraint on research automation, not proof that acceleration must wait for models to perform well on realistic tasks associated with ordinary jobs.












