AI NewsProductivityAnnouncement

METR says frontier AI models now handle a 40-hour task half the time

METR says GPT-6 Astra and Claude Opus 5.5 finish a 40-hour task half the time. Capabilities double every 3.6 months, by METR's own measure.

AI News

Editorial2 min read

LinkedInX
METR research report chart

Why it mattersA buyer gets a new figure to size an agent against. The figure is METR's own method, with wide uncertainty the report writes out.

METR, a non-profit model evaluator, published a new time-horizon report on 20 November 2026. It says frontier AI models now finish 40-hour software tasks half the time. A 40-hour task would take a working human a full week. METR says its measured doubling time is about 3.6 months. The 95 percent interval is 2.08 to 4.94 months. The report names GPT-6 Astra and Claude Opus 5.5 at that level. The test sets are METR's own TH1.1 and TH2.

What METR measured

Time horizon is the length of task, in human time, where a model finishes it half the time. The test sets are three. TH1.1 has 222 tasks pulled from older suites SWAA, HCAST and RE-Bench. TH2 has 36 tasks across 12 families, split into auto-scored and hand-scored. MirrorCode has 75 command-line software replication tasks. METR gives the human time for each task, then runs the model and marks pass or fail.

The TH2 auto-scored subset doubled faster, at 1.65 months. The 95 percent interval on that subset is 0.79 to 3.99 months, METR writes. On that subset alone, frontier models reach 27 hours of human time at the half-mark.

What METR says it does not know

The report is explicit about its own limits. METR says public evidence is too thin to rule out a large AI R&D speed-up soon. It says the team "remains significantly uncertain about central questions".

Only 9 percent of its human time estimates come from a real timed baseline, METR says. The rest are projected. The recruited baseliners may have been selected in a way that favours lower scores, METR writes. Models beat all three task baseliners in the rounds that were timed.

Another limit: some tasks look like the environments models are trained on in reinforcement learning. That could lift the score in a way the real world would not. METR says so.

A different gap in the model's judgement

The report also shows where frontier models fall off. METR asked a model to score another model's work. A frontier model matched METR's ratings when the task came from a weaker model. The model gave the work "inappropriately high scores" when the task came from one as strong or stronger, METR writes. A grader that cannot grade the strongest model fails first on the task that matters most.

What a buyer can do with it

A team buying agent work gets one new number to compare against. A product that runs for five hours of human time is far from the frontier. One that runs for 40 or more hours is at it, on METR's set. The buyer still has to run their own eval, because METR's tasks are not the buyer's tasks. The method the report uses is public, and its limits are stated.

Source

SourceMETR

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project