Has AI Progress Hit a Wall?

It is one of the most repeated claims in artificial intelligence this year: the models have plateaued, scaling is dead, the easy gains are gone. It gets louder every time a new release feels like a small step rather than a leap. So is it true? Mostly no, but the kernel of truth inside it is worth pulling out, because it is not the part most people mean.

Start with where the claim comes from. The line most often quoted is Ilya Sutskever's, that the internet is "the fossil fuel of AI," a finite reservoir of training data that we are burning through. That remark came from a talk in late 2024, not this year, and it was a point about data, not about intelligence hitting a ceiling. The distinction matters. Running low on fresh human text to train on is a real constraint. It is not the same as models being unable to get better.

What the evidence actually shows

The most credible source on the question is the International AI Safety Report 2026, a review drawing on hundreds of researchers. Its warning is economic, not technical. A frontier training run now costs on the order of $500 million in compute, it notes, with the next generation projected at $1 billion to $10 billion. If that spending stops paying off, the roughly fivefold annual growth in training compute could slow sharply, and progress would lean more on cleverer algorithms than on raw scale.

There is also real flattening on the old benchmarks. Models from OpenAI, Anthropic, Google and Meta show smaller jumps on the familiar English-language tests than their training budgets would suggest. But a flat benchmark is often a saturated one. When a test is nearly solved, there is little room left to climb, and the score stops being a useful measure of anything. That is a ceiling on the ruler, not on the model.

Where the gains moved

Look at the newer measures and the picture changes. The hard problems now are reasoning, long-horizon agent work, tool use and reliability, and those are still moving. A recent benchmark built around real economic tasks, which the best model scored just 12 percent on, is a good reminder of how much headroom remains once you stop grading work the models have already mastered. A plateau on a saturated quiz sits right next to a mountain of unsolved real work.

So the honest answer is layered. Scaling by brute force is getting expensive, and the returns on simply buying more compute are thinning. That is the true part. The claim that AI has stopped improving does not hold up, because the improvement has shifted to axes that the old tests were never built to see. It is worth being skeptical of leaderboard-topping announcements, as we have argued before, and skeptical too of the financial exuberance that assumes the curve only ever bends up. A wall is the wrong image. A fork in the road, where the cheap path gives out and the expensive one keeps going, is closer to the truth.