# METR: AI Time Horizons Double Every 3.6 Months, but Agents Lag Experts on Judgment

Mohit Tyagi · 2026-11-21 · Research

[https://timesof.ai/2026/11/metr-time-horizons-double-every-3-6-months-agents-lag-experts-on-judgment](https://timesof.ai/2026/11/metr-time-horizons-double-every-3-6-months-agents-lag-experts-on-judgment)

> The evaluation nonprofit METR said on November 20 that, across its older TH1.1 suite and a new, harder TH2 suite, the length of tasks frontier agents complete half the time is doubling roughly every 3.61 months (95% CI 2.08 to 4.94). Covering public models up to Claude Opus 5.5 and GPT-6 Astra, METR says current evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years, while finding agents have worse judgment than human experts on tasks without clear feedback. Most task times are estimates.

The evaluator best known for the "time horizon" chart has stopped arguing that today's models are too weak to worry about. In a capability update published on **November 20**, **METR** says public models now have strong enough coding and hacking skills that it "no longer" thinks it can argue risk is very low because of incapability on verifiable software tasks alone. Its new trend line keeps rising; the weak spot it found is judgment.

## The trend

METR evaluated public models up to **Claude Opus 5.5** and **GPT-6 Astra**. Across its older TH1.1 suite and a new, harder TH2 suite, time horizons, the length of task a model completes with 50% success measured in human time, are doubling roughly every **3.61 months** (95% confidence interval 2.08 to 4.94), or 110 days. That is close to the 105-day doubling METR previously reported on TH1.1 alone. Naively extrapolated, the trend reaches METR's longest score threshold, **200 hours**, in **July 2027**, the report says.

The ceiling is already close on older tests. METR says models saturated TH1.1, which includes tasks up to 30 hours, within a year of the metric's March 2025 launch, and current frontier models reach an estimated time horizon of up to **27 hours** on the verifiable part of TH2. On the three tasks for which METR tried to recruit human baselines, the models beat the humans it found.

The headline conclusion follows: METR says current public evidence is insufficient to rule out "drastic AI R&D acceleration" in the next couple of years, meaning capabilities improving so quickly that humans can no longer meaningfully understand, assess and react to catastrophic risks from new systems. It says it remains significantly uncertain about which capabilities are needed for that and how fast they are improving. METR adds that the combined trend is similar to the one OpenAI recently reported on [its own internal research tasks](https://timesof.ai/2026/11/openai-research-acceleration-agents-46-percent-two-to-four-hour-tasks-self-reported).

## Where agents fall short

On tasks that require choices without clear empirical feedback, METR finds AI agents appear to have **worse judgment than human experts**. In one task, model judges overrated outputs made by similarly strong or stronger models compared with an expert scorer.

MIT Technology Review, which interviewed outside researchers, reports that none of the AI-designed evaluations in one TH2 task was good enough for METR staff to consider it useful, and that models also fell short of human performance at scoring philosophical arguments. Opus 5.5 reached what METR calls "borderline above the bar" on the evaluation-design task, up from "clearly below the bar" for every model a year ago, the magazine reports. Sayash Kapoor, an incoming UC Berkeley professor, told it: "We're at the point where AI evaluations need to get more creative."

Our read: the doubling rate is the number that will travel, but the judgment gap is the one that matters for automating research, and it rests on the thinnest data in the report.

## Limits

- TH2 has 36 tasks in 12 families, and few are judgment-heavy, so the judgment findings rest on a small sample.
- Most TH2 task lengths are estimates rather than timed human attempts, MIT Technology Review reports, and in the philosophy task one person rated most arguments.
- The confidence interval on the doubling time is wide, and METR says the extrapolation is naive.
- METR has [deep-access evaluation arrangements with OpenAI](https://timesof.ai/2026/11/openai-embedded-assessments-apollo-metr-transluce-goodfire-deep-access); this report covers public models only.

## Sources

- [METR: AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure (November 20, 2026)](https://metr.org/blog/2026-11-20-q4-2026-capability-update/)
- [MIT Technology Review: The most misunderstood graph in AI just got a major update (November 20, 2026)](https://www.technologyreview.com/2026/11/20/1149399/the-most-misunderstood-graph-in-ai-just-got-a-major-update/)
