I want to pay for Opus where a review needs serious reasoning and send simple work to a cheaper model. That sounds obvious until you open a tool and leave the same model selected for everything. The subscription is available, the button is already set, and nobody asks why this particular job deserves that spend.
Anthropic launched Opus 4.5 in November 2025 at $5 per million input tokens and $25 per million output tokens. Looking back in October 2026, I care about that price because it opens up different ways of assigning work. My reference comes from a later split between the Sonnet and Opus families, without a record of the versions used.
In September 2026, I described my Claude setup: Sonnet implemented, Opus reviewed. That same month, I also asked whether simple tasks were going to the cheaper model. Wanting more capability during review fits perfectly well with being annoyed by waste during routine execution. I still think that combination makes sense.
Reserving a model for review means preparing its job. Handing over the entire repository and asking for a general opinion creates work for me too. I have to read the response, decide what applies, and separate actual problems from comments taking up space. An enormous review can be as inconvenient as a shallow one.
My rule is to give the reviewer a bounded change and its expected behavior. In a hypothetical example, a small edit might affect an important dependency. Diff size alone cannot decide how much reasoning it deserves. I want the analysis to examine consequences of that change and show where I can check each concern. Missing information should be named precisely.
Token prices belong in that decision alongside repeated work. A cheap call that keeps coming back consumes attention. An expensive call full of unrelated recommendations gains nothing from using the strongest model. I still need to measure that bill before calling any combination economical.
Sonnet implementing and Opus reviewing is an arrangement I can revisit as versions change. I want the role to stay explicit while the model choice remains open to revision. For the next review, I am starting with what it needs to discover. Then I can decide what that investigation deserves to cost.