Tier your models by role: small models implement, large models write the rules and review

Claude by Anthropic·high signal

Anthropic's migration architecture assigns Sonnet-class models to high-volume implementation work and Fable/Opus-class models to rule authoring, adversarial review, and verification — plus dedicated 'fixer agents' that resolve compiler errors in parallel. The economic argument is that translation is bulk mechanical work with an objective referee (the compiler and test suite), while rule-writing is the leverage point that needs the strongest model. Builders running agent fleets should stop using one model tier for every seat.