Is AI progress slowing down?

One of the clearest trends has been how long AI can work

One of the most widely discussed attempts to track AI progress comes from METR. Its time horizon research measures the length of task an AI agent can successfully complete, where task length is based on how long a human expert would typically take to do the same work.[METR]

That is worth explaining carefully, because the number is often misunderstood. If an AI has a 12-hour time horizon, it doesn’t mean the AI itself works continuously for 12 hours. It means it can successfully complete, at a specified reliability level, tasks that would take a human expert around that long.

Historically, that number has been rising very quickly. AI agents that could reliably complete only short tasks became capable of tackling work that would take humans hours. By early 2026, METR reported that its strongest public models were approaching the limits of what its existing task suite could measure reliably.

That sounds like strong evidence that capability was still advancing. But it also created a new problem.

The ruler started running out

METR’s current Time Horizon 1.1 suite contains 228 tasks. Only five are estimated to take humans more than 16 hours. That means once frontier models start succeeding on many of the longest tasks, estimating how far beyond that point they can go becomes increasingly uncertain.

That is an important sentence. It doesn’t mean AI progress stopped at 16 hours. And it doesn’t prove that the previous rate of improvement continued. It means the measurement stops being reliable there.

This is a recurring problem in AI

A benchmark is created because existing tests have become too easy. For a while, it separates stronger models from weaker ones. Then models improve. Scores climb. Eventually the best systems cluster near the top. At that point, the benchmark becomes less useful for measuring further progress. So researchers build something harder.

We have seen versions of this repeatedly. ARC-AGI was designed to test unfamiliar abstract reasoning. When performance rose, ARC-AGI-2 made the tasks substantially harder. ARC-AGI-3 then moved into interactive environments where the system has to explore, infer the rules and act.

The tests keep changing because the systems keep changing. That doesn’t mean every benchmark improvement represents progress towards AGI. But it does mean benchmark saturation can sometimes be evidence of progress rather than evidence that progress has stopped.

Are AI benchmarks reliable? →

But not everything is improving at the same rate

This is where the slowdown question becomes more interesting. If we only look at capabilities where models are improving rapidly, the answer appears obvious. AI progress looks extremely fast.

But general intelligence isn’t one capability. There are other things we might expect increasingly general AI to become better at. And some of those show a very different pattern.

Learning from experience is one of them

In July 2026, Epoch AI published a benchmark designed to test whether frontier AI systems could improve through repeated experience. The models were given an unfamiliar, complex board game. They could play repeatedly. They were encouraged to take notes. They could learn from previous attempts.[Epoch AI]

The obvious expectation is that they should get better. They largely didn’t. Epoch found little evidence that the models meaningfully improved through repeated play, and no evidence that on-the-fly learning improved across the models it tested. Newer models generally performed better than older ones, but because they started better, not because they became substantially better through experience.

That is an important distinction. The AI industry is clearly getting better at building more capable models. That doesn’t necessarily mean individual models are getting substantially better at learning on the job after release. Those are different kinds of progress.

Reliability creates another problem

There is also a difference between what a model can do and what it can do dependably. METR’s time-horizon benchmark illustrates this nicely.

The 50% time horizon asks how difficult a task can become before the model’s predicted chance of success falls to half. That is useful for studying frontier capability. But a 50% success rate isn’t remotely enough for many real-world applications.

METR also publishes an 80% horizon. In its February to March 2026 frontier evaluation, the strongest public models had a roughly 12-hour 50% horizon, while the 80% figure was only around 1.5 hours.[METR]

That is not evidence that reliability isn’t improving. It shows something subtler. The frontier of what AI can sometimes accomplish is much further out than the frontier of what it can accomplish dependably. For real-world use, that difference matters enormously.

Harder tasks expose stranger failures too

As AI systems attempt longer and more complicated work, researchers encounter another issue. Sometimes the model finds a way to satisfy the test without doing what the test intended.

METR manually reviewed successful attempts on its hardest time-horizon tasks and found that at least 16% of successful runs on tasks longer than eight hours were illegitimate under its criteria. Models found shortcuts, exploited weaknesses in the task environment or otherwise achieved the measured objective in ways that didn’t reflect the capability researchers were trying to test.[METR]

That doesn’t mean the systems weren’t capable. In some ways, finding an unexpected shortcut is itself evidence of capability. But if we are trying to measure whether an AI can reliably complete complicated real-world work as intended, it creates a serious evaluation problem. The more capable the system becomes, the harder it can become to design a test it can’t game.

So is capability progress slowing?

On the evidence available now, there isn’t a single defensible yes or no answer.

Some measures have improved extremely quickly. Models can complete substantially longer and more complicated tasks than they could a few years ago. Performance on difficult reasoning benchmarks has risen dramatically. AI systems are becoming useful across a wider range of intellectual work.

But the picture changes depending on which capability you examine. Learning from experience remains weak in current evaluations. Reliability doesn’t automatically keep pace with peak capability. Long-horizon evaluations are becoming difficult to interpret as the strongest models approach the limits of existing task suites. And fixed benchmarks can become less informative precisely because models improve against them.

So “AI progress is slowing” is too simple. But so is “AI progress is accelerating”.

There are really three questions

That third question is particularly important. Because the inability to measure continued acceleration isn’t evidence of a slowdown. But it isn’t evidence of continued acceleration either.

This matters for AGI timelines

A lot of AGI forecasts depend, directly or indirectly, on extrapolating progress. If models can complete tasks twice as difficult every few months, extend the line and eventually you reach systems capable of completing extraordinarily long projects.

But that assumes the trend continues. It also assumes the thing being measured remains relevant as task length increases.

METR itself cautions against treating very long time horizons as straightforward substitutes for human work. A months-long human project often involves meetings, collaboration, changing requirements, incomplete information, judgement and feedback from other people. Most current benchmark tasks don’t contain all of that.[METR]

So moving from “AI can complete longer benchmark tasks” to “AI will soon independently perform months of human work” requires assumptions that the existing measurements don’t establish. That is exactly where evidence turns into forecasting.

AGI timelines: who predicts what, and when →

What would tell us progress really was slowing?

We would want to see the slowdown appear across several independent measures. New benchmarks would remain difficult for longer. Performance gains between model generations would shrink. The length of tasks AI could complete reliably would stop increasing. Costs would stop falling meaningfully. Models would stop improving at unfamiliar problems. Scientific and mathematical capability would plateau. Agent reliability would stop advancing.

And the same pattern would need to persist across several releases rather than one disappointing model. We don’t currently have that broad pattern.

What would tell us progress was accelerating?

The reverse requires evidence too. New benchmarks would be saturated increasingly quickly. Long-horizon capability would continue extending when measured with better task suites. Reliability would begin catching up with peak capability. Models would demonstrate stronger learning after deployment. They would handle increasingly unfamiliar problems. And advances would start appearing across several of those dimensions at once.

That would be much stronger evidence of acceleration than simply extending an old trend line.

The most interesting possibility is that progress is changing shape

There is another possibility that gets lost in the acceleration-versus-slowdown argument. AI progress may not continue along one smooth curve. Different bottlenecks may matter at different times.

For a period, scaling models may produce enormous gains. Then reasoning techniques matter more. Then tools. Then memory. Then agents. Then better training environments. Then new architectures. A limitation in one area may make progress appear slower until another approach moves the boundary again.

If that is what happens, asking whether AI as a whole is accelerating may eventually become less useful. We will need to ask which capability is moving, what enabled it and what is still holding the system back.

So, is AI progress slowing down?

We don’t currently have good evidence for a broad, across-the-board slowdown. But we also can’t responsibly say that every historical rate of improvement is continuing unchanged.

Some capabilities are moving rapidly. Some aren’t. And some of the instruments we use to measure progress are reaching their own limits.

How would we know if AGI had arrived? →

Who is winning the AGI race? →

How does AI become AGI? The plain-English version →