Model Churn Exposure

Frontier assistants change month to month, so the real cost of standardising on one is not the seat price but the rework created each time a model tier shifts. Churn exposure is the share of client deliverables whose quality depends on a specific model version: prompt libraries, output templates, agent chains, and QA baselines all break or drift when the underlying model changes. Agencies should map each retainer deliverable to a churn tier (model-agnostic, model-tuned, model-locked) and price or buffer accordingly. The pattern is already visible: Gemini 4 Carbon went generally available on 10 November 2026 with a price doubling announced for 1 December, and Anthropic's Fable 5.2 changed coding behaviour based on how prompts framed code ownership. Both are silent rework events for any agency that hard-coded prompts to a prior version. Low-churn work stays portable; high-churn work needs version pinning, regression tests, and a named owner per client account.