A year ago, letting a model deliberate before answering felt like a trick. Now it's a default. Test-time compute — spending tokens on reasoning at inference rather than only on training — reshaped what models can do on hard problems. The interesting questions have moved on.
The curve is real, and it bends
Accuracy scales smoothly with how long a model is allowed to think, much as it once scaled with parameter count. But the curve has diminishing returns: the tenth sample helps less than the second, and past a point extra deliberation mostly burns money and latency for marginal gains.
The new question: how much is worth it?
Because thinking costs real tokens, the frontier problem of 2026 is allocation. Not "can it reason?" but "how much reasoning does this request deserve?" A trivial lookup shouldn't get a thousand tokens of deliberation; a hard proof should. Routing easy queries to short answers and hard ones to long ones is where cost and user experience are won or lost.
Compute is a budget now, not a switch. The skill is spending it where it changes the answer.
Adaptive reasoning
The most capable systems increasingly decide for themselves how long to think, calibrating effort to apparent difficulty. That turns a static setting into a dynamic one — and makes a cheap, accurate verifier even more valuable, since a good verifier multiplies the payoff of every sampled attempt.
For teams, the practical move is to measure reasoning spend per route and tune it deliberately. The wins now come less from a bigger model and more from spending inference compute intelligently.
