← Blog
Week of

Retrospective reconstruction · Published · Weekly archive

Route by the cost of being wrong

Model selection should account for consequence as well as request size.

Field NotesRetrospective

Retrospective field note for 2026-07-13, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.

A short request can be consequential. A long one can be an easily reversible draft. Token count alone is a poor guide to how much reasoning either deserves.

Classify the work by uncertainty and consequence. Is the evidence incomplete? Will the output change a persistent record? Can the result be checked cheaply before anyone relies on it? Those details should influence the route.

Keep that route inspectable. When a result fails, it helps to know which model handled it, what evidence it received and which check was supposed to catch the error. Otherwise every failure becomes a vague argument about model quality.

Escalation is allowed. A cheap first pass that identifies an unsupported case can save work. A cheap first pass that invents certainty sends the cost downstream.

The GPT-5.6 family broadens the menu. The product still has to make the choice. People should be able to specify speed, precision or a spending limit without learning a catalogue of model names before every task.

OpenAI introduced the GPT-5.6 family on 9 July. Source.

All posts Work with me