Claude Sonnet 5.5: calculate the real cost of repetitive tasks

In its announcement, Anthropic indicates a generation that is more than thirty percent faster and a cost up to thirty percent lower for most comparative tasks. The announced rates remain at 2 US dollars per million input tokens and 10 dollars for outputs, with a separate price for cached reads.

These gains are results from the provider under their evaluation conditions. They do not mean that all clients will save this percentage. The length of instructions, documents, tools, and corrections change the cost. Rates should be verified before estimating.

Sonnet can be evaluated on field extraction, case synthesis, or a preparatory response to a request. Define the inputs and the correct output. For a synthesis, require that important information is retained and that absences are not filled in by invention.

The test set should contain usual variations: long document, unreadable piece, multiple requests in a message, or contradictory instruction. An easy and well-written task does not necessarily represent daily work.

A quick response may require more verification. Conversely, a model that uses fewer steps can reduce cost while maintaining quality. Keep these dimensions separate to understand why a task works better.

First compare models with the same data and instructions. Then adjust the effort setting or architecture if necessary, keeping a record of the change. Do not compare an optimized process to an old version never adjusted without signaling it.

The cost must be linked to a stable definition of the accepted output. For an extraction, choose the essential fields; for a summary, the information to retain. Count separately the files processed automatically, corrected, and rejected. Savings calculated only on successful cases can hide the time spent on exceptions.

The Model evaluation must retain the batch volume and the share of accepted files. If the data or instructions change between two tests, indicate it in the comparison. Choosing a model for complex files is another task: the distribution of tasks.

A AI model integration requires cost limits, useful traces, and behavior when the supplier is unavailable. Plan for switching to another version, error tracking, and conditions for returning to the previous model.

Daillac did not test Sonnet 5.5 for this article. The news presents the announcement and an evaluation method, without recommending a universal migration. A local comparison must lead to a decision by use case, with the limits of its data visible.

Fictitious arithmetic example, without Daillac benchmark: a batch of one hundred requests consumes $4 in calls and requires twenty minutes of correction. Another consumes $3, but requires forty minutes. At a hypothetical internal cost of $30 per hour, the totals become $14 and $23. Tools and other possible costs must be added in the same currency. API invoice reduction is therefore not sufficient to choose the second process. The topic here is the savings from repetitive tasks, rather than a general model ranking.

Anthropic — Introducing Claude Sonnet 5.5. Sources consulted on October 1, 2026; the grids and decision examples are an editorial analysis by Daillac.