The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

AI Time Horizons Double Every 110 Days, METR Finds

By The Daily Byte ·21 Nov 2026 ·4 min read
A researcher studies lines of code and data on computer monitors in a dimly lit room (representative photo)

The length of the tasks that leading artificial intelligence models can handle on their own is still doubling roughly every 110 days, according to a new report from the AI evaluation nonprofit METR, published on Friday, 20 November. The update covers public frontier models through Anthropic's Opus 5.5 and OpenAI's GPT-6 Astra, and it examines the capabilities that matter most for automating AI research and development.

METR's "time horizon" measures the length of a task, scored by how long it would take a human to complete it, that a model can finish successfully about half the time. It is not a measure of how long the model itself works. The original plot, one of the most widely shared and misinterpreted charts in the AI debate, showed that length doubling about every four months. The new report says that pace has broadly held.

A harder test, the same trend

The update adds a second suite, TH2, with 36 tasks in 12 families: eight families with well-defined, automatically scored objectives, covering optimisation, machine learning, puzzles and games, and four families covering conceptual reasoning, judgment and open-ended research, where success has no cheap automatic check. METR typically evaluated 15 models on the two suites combined.

Across the combined TH1.1 and TH2 evaluations, the estimated doubling time for the 50 percent time horizon is about 110 days, with a 95 percent confidence interval of 2.08 to 4.94 months, close to the previous 105-day trend for TH1.1 alone. On the verifiable subset of TH2, the estimated doubling time falls to 1.65 months, but METR cautions the noisier result could reflect task composition or selection rather than broad acceleration.

On verifiable TH2 tasks, METR estimates the current public frontier models' time horizons at up to 27 hours. On each of the three tasks with formal human baselines, the best model outperformed every recruited human baseliner. Across 15 Al Zimmermann optimisation contests, the strongest models placed at the 96th to 98th percentile of human participants in the median contest, although the comparison is uneven: agents had 48 hours on a 28-core machine, while humans had about three months.

Judgment still lags behind coding

The hard-to-verify tasks tell a different story. In one, models had to design an informative new evaluation task for METR's own suite. None of the AI-designed evaluations was good enough for METR employees to consider useful for their research, although the newest models, such as Opus 5.5, reached what METR calls "borderline above the bar" performance, up from "clearly below the bar" a year ago. In another task, models had to score philosophical arguments, a test of judgment on which they also fell short of human performance.

The gap matters because the tasks that are easiest to train and measure, coding and mathematics, are the ones where success can be checked automatically. "We're at the point where AI evaluations need to get more creative," Sayash Kapoor, a computer scientist and incoming professor at the University of California, Berkeley, told MIT Technology Review. "Trying to push the envelope of the space of evaluations is a welcome change."

What it does not settle

METR is candid about the limits of its own measurement. In the new report the human completion times are mostly estimates, because paying people for tasks that might take hundreds of hours is prohibitively expensive: direct human baselines informed only 9 percent of the estimates, with researcher estimates used for 31 of 34 tasks. Results above roughly 16 human-task-hours remain unreliable. "By their nature, hard-to-verify or judgment-laden tasks are much harder to study, and especially to quantify," Matt Putz, a member of METR's technical staff, said. "Ultimately, I don't think we'll be able to establish trends as clean and robust as for the easy-to-specify [tasks]."

The report also notes that current public evidence is insufficient to rule out drastic acceleration of AI research and development in the next couple of years, meaning capabilities improving so quickly that humans can no longer meaningfully understand, assess and react to catastrophic risks. Weaker judgment is a plausible constraint on research automation, METR says, not proof that acceleration must wait. The update follows a separate, heavily redacted METR report from earlier in November which claimed AI might be speeding up the pace of AI progress at Anthropic by a factor of two.

For now, the picture is split: on work with a clear right answer, frontier models keep improving at an exponential pace. On the open-ended judgment that research depends on, they are improving too, but from further behind, and the evidence is thinner. "The improvement we observe, I think that does tell us something for the current moment," said Najoung Kim, a linguist and computer scientist at Boston University. "I'm not sure just yet how to think about how it would extrapolate."