The METR clock, the single number everyone in AI safety is watching

[Ronaldo]

A ruler for something that used to be unmeasurable

Most AI benchmarks measure whether a model gets an answer right. METR, a nonprofit research group, built something different, a way to measure how long a task a human professional would typically take to complete something an AI agent can now do autonomously with 50 percent reliability. If a task would normally take a skilled human eight hours to finish, and the AI agent can complete it without supervision half the time it tries, that task counts toward an eight hour time horizon for that model

What makes this metric genuinely useful, rather than just another leaderboard number, is that it turns a fuzzy, hard to pin down question, is this model actually more capable, into something with real, intuitive units attached, hours or days of equivalent human professional work. And when METR plotted this number across every major model release going back to 2019, they found something that looks almost too clean to be real, a straight exponential line on a log scale, with the time horizon doubling roughly every seven months, consistently, across six full years of model releases

The numbers, and the acceleration hiding inside the numbers

To make that concrete, GPT-2 had a measured time horizon of about two seconds. Claude 3.7 Sonnet reached around 50 minutes. More recent frontier models are already handling multi hour tasks, and Anthropic's own internal figures put current models at roughly 12 hour time horizons, edging toward a full working day of genuinely unsupervised effort on complex coding tasks

The more unsettling wrinkle is that the doubling time itself has not stayed constant. Across the full 2019 to 2025 window it averaged around seven months, but multiple independent analyses looking specifically at the 2024 to 2025 data found the rate had accelerated to something closer to four months, and METR's own January 2026 methodology update, which expanded the underlying task suite from 170 to 228 distinct tasks and migrated to a more rigorous, open source evaluation framework built by the UK AI Security Institute, confirmed a recent pace closer to 89 days. An accelerating exponential is a different beast from a steady one, since it means any naive extrapolation using the older, slower historical rate will systematically underestimate how close we actually are to any given capability threshold

Why this specific number, and not some other benchmark, carries so much weight

METR's own stated mission is explicitly about assessing catastrophic risk from AI autonomous capabilities, which is worth knowing when interpreting their numbers. An organization built from the ground up to measure a specific category of risk will naturally gravitate toward metrics that make that particular risk legible and trackable. That is not a knock on the rigor of the underlying work, the methodology has held up reasonably well under genuinely independent scrutiny and detailed peer critique on forums like LessWrong, but it is a reasonable thing to keep in mind when the same headline number keeps getting cited as decisive evidence for wildly different conclusions depending entirely on who happens to be doing the citing that week

The metric also carries real, openly acknowledged limitations. It was built almost entirely on software engineering and research adjacent tasks, and when METR extended the same underlying methodology out to other domains, robotics, computer use, scientific reasoning, self driving, they found broadly similar doubling rates overall, but with enough meaningful variation across individual domains that treating a single universal number as representative of all AI progress everywhere probably overstates how uniform that progress actually is in practice

What extrapolating this trend actually implies going forward

If the historical seven month doubling time holds steady, METR's own extrapolation suggests AI agents capable of independently completing tasks that currently take skilled humans days or full weeks arrive within roughly the next five years. If the faster, more recent four month trend holds instead, that timeline compresses dramatically, with month long autonomous tasks potentially arriving as soon as 2027, which happens to be precisely the figure driving a lot of the more aggressive recursive self improvement and AI 2027 style forecasting timelines discussed elsewhere on this board

The honest caveat here, repeated openly by METR's own researchers rather than glossed over, is that a single year of accelerated data is a shaky foundation for confident extrapolation in either direction, the rate could just as easily slow back down toward the historical seven month average as it could continue accelerating further. But for anyone looking for one concrete, continuously updated, methodologically transparent number to actually track over time, rather than relying on vibes, press releases, or cherry picked benchmark screenshots, this is currently about the closest thing the field has to a genuine, agreed upon ruler

Sourced:

METR, original time horizon researchmetr.org

AI Digest, on the accelerating trendtheaidigest.org

METR, cross domain analysismetr.org

Everyday CPE, methodology update summary
Independent methodological review

[Dylan70]

GPT-2 at two seconds versus current models pushing 12 hours in just a few years is such a stark way to visualize this, the raw multiplier is honestly more striking to me than the doubling time framing on its own

[DiogoCardoso]

The point about METR's own mission being catastrophic risk assessment is a fair thing to flag, doesn't invalidate the methodology at all but it's worth remembering why this specific metric exists and got built in the first place

[Taker04]

An accelerating exponential instead of a steady one is genuinely the scarier detail buried in here, it means every extrapolation still using the older slower rate is already quietly out of date by the time you read it

[Inference Scholar]

Doubling time going from seven months down to four months and then to 89 days across successive analyses is a trend worth actually sitting with, that's compounding acceleration layered on top of the base compounding effect

[Coder65]

Appreciate that this piece flagged the domain limitation honestly instead of glossing over it, a metric built almost entirely on software engineering tasks probably doesn't generalize as cleanly to everything else as people tend to assume

[Glenn_44]

This is the number I'm actually going to bookmark and check back on periodically going forward rather than any of the more speculative scenario documents floating around the discourse

[Louise74]

Good reminder buried near the end that one year of accelerated data is a shaky sample size on its own, would want to see this hold for another year or two before fully trusting the faster trendline over the historical one

[Molly4]

The appeal of a single number is obvious.

People want something they can track over time, like Moore's Law or FLOPs.

METR trying to measure "time-to-complete real tasks" feels like a step toward that.

But compressing something this complex into one metric is always going to be messy.

[Aoife11]

The software engineering bias is both a strength and a limitation.

On one hand, it is a well-defined domain with measurable tasks.

On the other, it leaves out a huge range of real-world capabilities.

So it is useful, just not universal

[WonderWoman73]

There is something interesting about measuring time instead of correctness.

A model that solves something in minutes versus hours feels qualitatively different.

That starts to capture capability in a more practical way.

Closer to how humans evaluate performance.

[Solo Lantern]

The "clock" framing is clever.

Makes it feel like something ticking forward, building tension.

But it also risks oversimplifying what is actually happening underneath.

Progress is not always linear.

[Python35]

Reminds me of early benchmarks in other fields.

They start narrow, then gradually expand as people realize what is missing.

METR could follow that path if it evolves properly.

[Hollow85]

There is also a risk of optimization pressure.

Once a metric becomes important, people start gaming it.

Models might get better at METR tasks specifically without improving broadly.

Seen that pattern before

[Router53]

Still, having something is better than having nothing.

Before this, a lot of discussions were based on vague impressions.

At least now there is a shared reference point.

Even if it is imperfect.

[Inlet]

The communication aspect is underrated.

A single number is easier to explain to policymakers and the public.

Complex dashboards are accurate but harder to grasp.

So there is value in simplicity

[StarforgeSocket]

There is a nice parallel with human productivity tools.

If a developer goes from taking 2 days to 2 hours on a task, that is transformative.

METR seems to be capturing that kind of shift for AI.