Why we don't ship GPT-4.1 Nano as the default model

The task that breaks small models is the compound one: read a long email in one language, translate it, draft a sensible reply, and pick the correct attachment — all in a single turn. Nano drops steps. It mistranslates names. It attaches the wrong PDF.

So our default is Sonnet 4.6, with GLM-5.1 for cheap routing and Opus on demand for the hard cases. When a tenant hits their soft cap we throttle to the cheaper model rather than locking the channel — the parent on the other side is never stranded mid-cycle.