METR: Claude, GPT-6 Task Horizons Double Every 3.6 Months

The length of task that frontier AI agents can finish is doubling roughly every 3.6 months, METR said on Friday. Its new test suite covers models up to Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Astra.

The AI evaluation nonprofit also found a gap. On tasks without clear feedback, “AI agents appear to have worse judgment than human experts”, METR wrote.

Business Pill: The Cost of Checking

An evergreen Business Engineer explainer on the cost of checking work. It teaches the general idea only and does not discuss METR’s results.

The key insight: METR says the length of task frontier AI agents can finish at 50% success is doubling about every 110 days, reaching up to 27 hours on its verifiable tasks, while on judgment-heavy tasks without clear feedback METR finds the agents appear to trail human experts.

What METR Measured

METR’s time horizon is the length of task, measured in human working time, that a model completes with 50% success.

When METR first published the metric in March 2025, the best model managed about 1 hour, Sonnet 3.7. Within a year, models had saturated its original suite, which runs to 30-hour tasks.

So METR built a harder suite, TH2: 36 tasks in 12 families. Eight families are verifiable, covering optimization, machine learning, puzzles and games. Four cover conceptual reasoning, judgment and hard-to-verify research.

It ran 15 models through both suites, from GPT-4o and o1 to Opus 5.5 and GPT-6 Astra.

METR’s doubling time for the 50% time horizon, before and after it added the harder TH2 tasks.
METR’s doubling time for the 50% time horizon, before and after it added the harder TH2 tasks.

Doubling Every 110 Days

Across both suites, time horizons are doubling every 3.61 months, or 110 days, with a 95% confidence interval of 2.08 to 4.94 months. The earlier trend on the original suite was 105 days.

On the verifiable part of TH2, current public frontier models reach an estimated time horizon of up to 27 hours, METR said.

METR notes that “naive extrapolation” of the trend would put models at a 50% time horizon of 200 hours, the longest threshold in its tasks, in July 2027. That is METR’s extrapolation, not a forecast.

METR Figure 10: 50% time horizon by model release date on a log scale, from GPT-4o to GPT-6 Astra, doubling roughly every 3.6 months
Source: METR, Figure 10, 20 Nov 2026 (CC-BY). Frontier doubling time 3.61 months, 95% CI 2.08–4.94.

Where Models Beat People

In 15 Al Zimmermann programming contests, multi-month optimization competitions, the median placement of the strongest models was the 96th to 98th percentile of human entrants, METR found. They were “generally unable to beat the very top performing humans”.

On MirrorCode, Opus 5.5 solves 77% of medium to large programs that would take human experts weeks to months, METR wrote. When METR recruited its own human baseliners, it “typically failed to find humans that outperform the current frontier models”.

Where They Fall Short

The judgment gap shows up when feedback is slow. In one task, model judges overestimated an output’s quality when a similarly strong or stronger model had produced it, compared with an expert scorer.

Asked to design METR-style tasks, agents often still produced hill-climbable machine-learning tasks, even when steered toward judgment-loaded ones. In a novel board game, frontier models were worse than humans at improving over time.

Models also did worse on tasks METR estimates are costlier to turn into training environments. A doubling of that “Lifecycle Cost” is very roughly equivalent to halving an agent’s time horizon, though progress ran at similar rates on cheap and costly tasks.

The Structural Read

As we read METR’s post, the headline is less the speed than the split: fast progress where answers can be checked, a lag where they cannot.

As we read the Lifecycle Cost finding, the tasks that are hardest to build training environments for are also where models score worst, which ties evaluation cost to capability.

METR, 20 Nov 2026

“We think that current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years.”

Three Implications

FOR AI LABS METR’s frontier models placed at the 96th to 98th percentile in multi-month programming contests.

FOR EVALUATORS METR warns its containerized evals could overestimate real-world capabilities.

FOR BUYERS OF AGENTS METR found model judges overrated outputs from similarly strong models.

What Is Not Established

METR says it remains “significantly uncertain” about which capabilities matter for drastic acceleration of AI research, and how fast they are improving.

It also warns that its easy-to-run evals “could overestimate real-world capabilities” because similar training environments have become common. METR concludes that public evidence is “insufficient to rule out drastic AI R&D acceleration in the next couple of years”.

Separately today, an Anthropic mathematician claimed a proof of the Riemann Hypothesis made with Claude; that claim has not been verified.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of where AI capability turns into business leverage.

Read the Map of AI →

The Bottom Line

METR’s frontier models now clear up to 27 hours of verifiable work at 50% success, and the horizon doubles every 110 days, but METR finds their judgment appears to trail human experts on tasks without clear feedback, and it says the models may be underelicited there.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. All capability figures are METR’s, from its 20 November 2026 post by Megan Kinniment, Beth Barnes and colleagues; the chart reproduced below is METR’s Figure 10, published under CC-BY. The 200-hour, July 2027 figure is METR’s own naive extrapolation, which METR does not present as a forecast. Model names are as METR reports them.

Sources: METR, ‘AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure’, 20 Nov 2026 · METR, Task-Completion Time Horizons of Frontier AI Models · Techmeme, 20 Nov 2026 (MIT Technology Review, The Information)

Scroll to Top
FourWeekMBA