METR finds AI agent time horizons doubling every 3.6 months

The evaluation nonprofit's latest update says frontier models now beat its recruited experts on coding and optimization tasks but still show weaker judgment where results are hard to check.

Illustration of a robot climbing a staircase of glowing hourglasses toward a large wall clock while a human researcher with a clipboard and measuring tape looks on. Image: AgenticFeed illustration

The length of tasks that AI agents can complete on their own is doubling roughly every 3.6 months, according to a capability update published November 20 by METR, the nonprofit that runs independent evaluations of frontier models. The group also said current public evidence is "insufficient to rule out drastic AI R&D acceleration in the next couple of years."

METR measures a model's "time horizon": the length of a task, measured by how long a skilled human expert needs to finish it, that an agent completes with 50% reliability. Pooling its older Time Horizon 1.1 suite with a new, harder suite called TH2, the researchers estimated a doubling time of 3.61 months, or 110 days, with a 95% confidence interval of 2.08 to 4.94 months. That is close to the 105-day trend METR previously found on Time Horizon 1.1 alone. The update covers public models up to Anthropic's Opus 5.5 and OpenAI's GPT-6 Astra.

Harder tasks, pricier baselines

METR wrote that when it introduced the metric in March 2025, the best model had a time horizon of around one hour. Within a year, models had saturated Time Horizon 1.1, whose 222 tasks run up to 30 hours of human work.

TH2 adds 36 tasks in 12 families, according to METR. Eight families are "verifiable," covering optimization, machine learning, puzzles and games scored automatically. The other four involve conceptual reasoning, judgment and open-ended research with no cheap check for success. The group also tested a separate set of 75 software replication tasks called MirrorCode.

On the verifiable part of TH2, current public frontier models have an estimated time horizon of up to 27 hours, the report says. Fitted on that subset alone, the doubling time drops to 1.65 months, though METR called its data noisy and listed explanations other than a real speedup.

Finding humans who can still beat the models has become expensive. METR said it recruited specialists, including a former International Grandmaster on Codeforces, paid $200 to $350 an hour and spent roughly $90,000 on seven attempted baselines. Frontier models beat those baseliners on all three tasks where METR tried to collect human results.

Judgment remains the gap

On tasks that require choices without fast feedback, the report says AI agents "appear to have worse judgment than human experts." In one task, model judges overrated the quality of work produced by a similarly strong or stronger model when compared with an expert human scorer.

METR also tested agent swarms, groups of agents working on a task together, against a best-of-k baseline that keeps the best of several independent runs. Swarms helped on tasks with a clear metric to optimize but had unclear or negative effects elsewhere, and large swarms hurt performance when agents had to build their own proxy metric. The swarm experiments used about $500,000 in tokens, according to the report.

METR said a naive extrapolation of the combined trend would put models at a 200-hour time horizon, the longest threshold in METR's task set, in July 2027. The authors cautioned that easy-to-run evaluations may overestimate real-world ability because similar environments are increasingly common in model training.

The researchers wrote that public models now perform strongly enough on coding and hacking that METR can no longer argue risk is very low based on weak performance on verifiable software tasks alone. They said the doubling time on the combined suite is similar to a trend OpenAI recently reported on its internal research tasks.

METR tracks progress with a composite measure it calls the METR Capability Index. The group said it does not see strong acceleration or deceleration across the public frontier on that index, though it found uncertain signs of trend changes at individual developers.

Source: METR, Megan Kinniment, Beth Barnes and others, “AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 30 Nov, 2026

AWS re:Invent 2026 in Las Vegas

AWS’s annual conference runs Nov. 30 to Dec. 4. Matt Garman’s CEO keynote is on Dec. 1, and the catalog lists more than 2,200 sessions.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.