Most AI models think for a fixed amount of time. Ask an easy question or an impossibly hard one, and the model spends exactly the same effort on both before it answers. It is a strange way to build something meant to reason. No human works that way — we glance at "2 plus 2" and labor over a hard puzzle. So we built models that decide for themselves how long to think, and then we asked an obvious, slightly reckless question: what happens if we just let one keep going?
Our models think in rounds. Each round is a step of internal computation, and the model runs a little internal race to decide when it has thought enough and should commit to an answer. Normally that race ends quickly on easy inputs and runs longer on hard ones — that is the whole point. But you can override it. You can tell the model to ignore its own stopping instinct and keep looping.
So we did. We took problems harder than anything the model had trained on, removed the brakes, and watched.
The interesting part is that thinking longer actually worked — up to a point. On reasoning problems much longer than it had ever seen, a model that was allowed to keep looping solved them, while the same model forced to stop early failed. The extra rounds were not wasted motion; they were real computation, and they let the model reach answers that were otherwise out of reach. In our tests, a model trained only on short problems could solve chains more than twice as long, simply by being given the time to work through them.
Given more time to think, the model solved problems it was never trained to handle. The extra rounds were doing real work.
But "forever" is not the same as "longer." Past a certain point, more thinking stopped helping. The model had already reached its answer; additional rounds just churned. This is exactly why the stopping mechanism matters. A model that thinks forever is not wiser — it is wasteful. The skill is not thinking longer; it is knowing when you are done.
This little experiment points at something we believe is important. The dominant way to make AI smarter right now is to make it bigger — more parameters, baked in at training time. But here was a model getting measurably better at harder problems without changing a single weight, just by spending more thought at the moment of answering. Capability was not only in the size of the model; it was in how it used its time.
That is a different and cheaper lever than raw scale, and it is one we keep pulling. A model that spends its thinking where it is needed — more on the hard, less on the easy, and stops when it is done — is doing more with less. Which, around here, is the whole idea.
The honest caveat, as always: this was measured on structured reasoning tasks, not the open-ended messiness of real language, and "think longer" only helps when the underlying steps actually compose. But as a glimpse of what adaptive computation can do, it is one of our favorite results — a small model punching above its size, simply because we let it take its time.