Claude Opus 5.5 Cuts Costs With 40% Fewer Tokens

Anthropic released Opus 5.5 this week, the first model in its Claude 5.5 line, and the headline numbers are about money as much as intelligence. The company says the model runs more than 30% faster than Opus 5, costs less per token, and finishes typical workloads using roughly 40% fewer tokens overall. Put together, Anthropic is claiming a real drop in what it costs to run advanced reasoning at scale.

That matters because token costs are what determine whether an AI feature is cheap to ship or quietly expensive to maintain once usage climbs. A model that reasons well but burns through tokens fast can end up costing more than a slower, cheaper one. Opus 5.5 is being pitched as solving that tradeoff: sharper writing, quicker replies, and a smaller bill at the end of the month.

Anthropic also changed how the model handles reasoning by default, a shift that affects both how developers configure it and how it performs out of the box. Before getting into benchmark results and hands-on tests, it's worth looking at exactly what changed under the hood, starting with how Opus 5.5 decides when and how hard to think.