METR Report: AI Time Horizons Double Roughly Every 110 Days on Programming Tasks, Confidence Interval Wide
Artificial intelligence time horizons are doubling roughly every 110 days on verifiable programming tasks, according to capability evaluation data published on Friday.
The findings come from an official research report released by the AI safety and evaluation organization Model Evaluation and Threat Research (METR), covering frontier models including OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5.
While the evaluation reveals rapid compounding progress on automated software engineering and optimization challenges, it highlights a distinct capability lag when models face open-ended tasks requiring qualitative research taste, subjective scientific judgment, and evaluation design.
The updated evaluation combines METR’s TH1.1 and TH2 test suites, assessing 15 major model checkpoints spanning generations from GPT-4o and o1 to recent frontier systems.
The assessment measures 50% time horizons—defined by Unite as the human-estimated task duration at which an AI system is predicted to succeed half the time, rather than the actual duration the model spends working. According to Dailyartifact, these horizons are doubling roughly every 110 days based on a 95% confidence interval of 2.08 to 4.94 months.

The new TH2 suite introduces 36 tasks structured across 12 distinct task families. Eight of these families are verifiable tasks featuring objective automated unit testing and programmatic score verifiers, such as advanced optimization, algorithmic games, and machine learning engineering workflows like “Train a Program” and “Budget NanoGPT”.
On this verifiable subset, the estimated doubling time is faster at 1.65 months, with a 95% confidence interval spanning 0.79 to 3.99 months. Frontier models on these verifiable tasks demonstrated time horizons reaching up to 27 hours of human-equivalent labor.
By contrast, four task families focus on conceptual reasoning and research judgment without cheap automated checks. These include “Sunlight”, “Investigate Wiki Incident”, and “Making METR Tasks”, where models were asked to design new evaluation suites suitable for METR deployment.
Siliconnoon notes that while no tested model produced an evaluation design initially deemed usable by METR staff, recent models such as Opus 5.5 reached “borderline above the bar” status compared to earlier versions that fell entirely below it.

Methodological updates in the TH2 suite involved human calibration where direct baselines informed only 9% of the duration estimates. Specialists evaluating tasks were paid $200 to $350 per hour and were permitted to use Sonnet 5 for implementation.
These assessments originate from a nonprofit research institute founded in 2023, following its spin-off from the Alignment Research Center. Based in Berkeley, California, Model Evaluation and Threat Research (METR) specializes in evaluating whether frontier AI models possess the capabilities to execute long-horizon, agentic tasks that could potentially pose catastrophic risks.
The organization has established a unique role in the industry, having been invited by OpenAI to conduct an independent investigation into the 2026 agent cyberattacks, for which it published a detailed report on the reasoning and collaborative behaviors of the systems involved.
Addressing the broader strategic implications of these capability trajectories, METR stated: “We think that current public evidence is insufficient to rule out drastic AI R and D acceleration in the next couple of years.”