A useful division of work needs to remain open to review. Previous use gives no model a permanent job. When a new version arrives or the assignment changes, I want to check whether the setup still helps produce an acceptable change without leaving the rest to the reviewer.
On September 28, 2026, I asked whether Sonnet would be better for implementation. Two weeks earlier, on September 15, I had described using Sonnet to do the work and Opus to review it. The question remained open despite that practice. The records name model families and establish no use of a particular version.
Anthropic released Sonnet 5.5 on that September 28. Pricing stayed at $2 for input and $10 for output per million tokens. The company attributed savings of up to 30% per task to lower consumption. That is a concrete reason to reconsider frequent implementation work. The percentage remains the announced estimate; I have no measured personal savings to add.
That day, I also asked for an additional request to be delegated in parallel to preserve ongoing work. That situation explains why the handoff interests me so much. An executor needs to understand where its responsibility ends and what it should return. If it depends on an unresolved decision elsewhere, adding another model may only produce another version of the same uncertainty.
The implementation I want connects requested behavior with the change and evidence that it works. Speed helps when it produces a usable first version and allows short correction cycles. A diff that is hard to understand collects that time again during review. I need to locate what changed and why without reconstructing the entire conversation preceding the code.
To compare the division, I will use bounded tasks varied enough to expose real difficulties. Review needs to verify the expected result. Another agent's approval alone cannot settle that check. Total effort includes work returned for correction and information the reviewer had to seek because it was missing from the initial delivery.
My decision is to retain Sonnet as a familiar starting point while reopening the choice around 5.5, without announcing a winner beforehand. I want to record cost and the route to a reviewable change for each comparison. If another setup delivers that with less work returned to me, I will switch. The implementation job continues to depend on the next deliverable.