← Blog
Week of

Retrospective reconstruction · Published · Weekly archive

Cheap calls. Expensive afternoon.

The API bill doesn't include the time spent fixing the output.

Field NotesRetrospective

Retrospective field note for 2025-10-27. Reconstructed in November 2026 from dated sources, not originally published on the archive date.

The API bill doesn’t include the time spent fixing the output.

After the Haiku 4.5 release, it’s worth comparing routes by the cost of a completed task rather than the price of a call. Otherwise the dashboard can reward a model that creates more work for its user.

Track retries, failed validation and escalation. Keep the failure examples too. An average can hide the one category of input that repeatedly goes wrong.

This doesn’t mean always choosing the larger model. Some mistakes are cheap to catch and repair. Others change external state or interrupt a person. The route should reflect that difference.

Before upgrading, see whether the task itself can be made clearer. A better schema might remove an entire class of errors. A deterministic check might settle something a second model would merely debate. There is no reason to pay for reasoning where a simple test gives an exact answer.

I’d want the comparison to end with a practical choice for a particular workflow, not a declaration that one model is best. Include the cleanup. If the supposed saving appears as another hour of supervision, the cost hasn’t disappeared. It has moved to a less convenient column.

Sources

All posts Work with me