Model Deprecation: What Happens When the Model You Built On Is Retired

Every model has a retirement date, whether or not it has been announced yet. On the API the process is orderly: the vendor marks a model as deprecated, publishes a date usually several months out, names a successor, and on the date requests to the old model start failing. In a consumer app the process is invisible: the model behind the product changes when the vendor decides, and a workflow that depended on last month's behaviour finds this month's different. The first is a scheduled event that many teams treat as a surprise. The second is a surprise that no amount of reading can schedule, which is why anything that matters belongs on the API.

The damage from either is regression, and it is worth understanding why. A prompt is tuned to a model's habits - how it structures a list, how literally it follows an instruction, what it does with an ambiguous case - whether or not the person who wrote it meant to tune it. A successor has different habits. The regression is rarely total and rarely nothing, and it lands in the cases nobody thought to test, because those are the cases where habits show. Cost and latency move too, and a successor with a lower per-token price can cost more per task if it writes longer or reasons harder.

The defences are unglamorous and they work. Build on a dated snapshot rather than the floating alias, so the model under you changes only when you change it, and read the model name from configuration so that changing it is a config edit rather than a code search. Keep an evaluation set: twenty to fifty real cases from your own workload with the outputs you consider correct, and run any candidate model against it before switching. That set is what turns a deprecation from an outage into a comparison, and it takes an afternoon to build. Watch the vendor's deprecation page monthly, or subscribe to it, and when you adopt a model put its likely lifetime and a migration window on the calendar. A retirement with six months of notice is a task; the same retirement discovered on the day of failure is an incident, and the only difference is attention.