Anthropic has just presented Claude Opus 5.5, launched on 22 September, as a model less costly to run than Opus 5. The lab announces savings of around 40% on its typical workloads, at default settings. This usage estimate says nothing about a uniform drop in the price of each token, and the nuance changes the reading.
Two savings overlap. The listed rates are 4 dollars per million input tokens and 20 per million output tokens, against 5 and 25 for Opus 5, that is 20% less per unit. Cache reads, which allow an already processed context to be reused, fall from 0.50 to 0.20 dollars per million. Anthropic adds that the model consumes fewer tokens per task, and it is this economy of tokens, more than the rate, that carries the overall gain.
When a Maison compares tools for analysing its archives or preparing a collection dossier, knowing where the saving comes from matters. A lighter bill may come down to the rate, to the amount of work accomplished or to a better-reused context. Each of these causes commits the organisation differently.
The developer also puts forward a reservation rarely seen on this kind of page: at this level of capability, score gaps between models are less informative about real differences. If the tables no longer separate the models, the comparison has to be made elsewhere, on the Maison’s own ground.
A useful evaluation would start from a few representative dossiers, each accompanied by an expected result and the list of errors that would make it unacceptable. One would then count the rework, the review time and the missing sources, in addition to the amount billed by the provider. The question then becomes precise: which tool produces a dossier the team can actually use, and at what full cost? A cheap result that takes long to verify does not remove the expense. It shifts it to the most experienced people, those whose time is most expensive.
The announcement therefore opens a possibility of reducing technical spending. Its value for a Maison remains to be established on its own work, with its requirements of proof, confidentiality and transmission.
Cette publication est également disponible en :
