一番賢いモデルではなく、仕事に合うモデルを選ぶ

Choosing a model is a tradeoff between intelligence, speed, and cost. Anthropic raises capability with two levers. Train time compute improves the model's base capability during training, while test time compute is extra compute used at inference time, often by letting the model spend more tokens thinking, calling tools, or working through a task before it answers. More tokens tend to raise quality and accuracy on hard problems, but they also raise latency and cost, and the gains often show diminishing returns at the highest effort. So the goal is not always the smartest model. The best way to choose is not to guess but to build hard eval sets and measure each option. A useful rule of thumb from the docs: for complex reasoning or coding, start with a larger model; for simple, high-volume, or low-latency tasks, a smaller one is often enough.