Anthropic keeps Sonnet’s price unchanged but announces lower consumption. For a Maison, the comparison should rest on an accepted file, corrections included.
The listed price does not fall: Claude Sonnet 5.5 is still billed at 2 dollars per million input tokens and 10 dollars per million output tokens. Yet Anthropic announces up to 30% lower cost per task in its tests. Presented on 28 September 2026, the model is said to use fewer tokens to accomplish the same work. The provider also claims text generation that is more than 30% faster.
These two measures should not be conflated. Generation speed concerns a single step; the cost of a file also includes research, tool calls and revisions. Anthropic positions Sonnet on well-defined tasks, while reserving for Opus 5.5 its advantage on open-ended work that demands sustained judgment. The reported performance remains that of the laboratory and of the users it cites.
A trial based on the Maison’s requirements
To prepare a collection review, a team could give the candidate a file it already knows: the same documents, the same questions, the same admissibility criteria. It would then compare omissions, attributions and the time spent on corrections. This protocol is a proposed evaluation, not a result obtained at a client.
The effort setting also deserves a place in the written account. Two trials carrying the same model name may use different durations and budgets. Recording this setting would make it possible to reproduce the test when the corpus changes.
The announced saving becomes worthwhile if it leaves the expected documentary quality intact. A model that delivers a text faster but forces every reference to be redone consumes precisely the knowledge whose time a Maison seeks to preserve: that of the people able to recognise an error in its archives
Cette publication est également disponible en :
