Match the model to the task, don’t default to the largest frontier model for everything.
Easy to implement
8 High impact score
A model many times larger than necessary burns proportionally more energy on every single inference call, whether it’s summarising one sentence or drafting a full report. Reserve frontier-scale models for tasks that genuinely need their reasoning depth, and route simpler jobs, classification, extraction, formatting, routing, lookups: to smaller, cheaper, faster models, or to a non-LLM approach entirely where one already does the job (see 8.6‘s companion principle: not every task needs AI at all). This is the same right-sizing discipline as infrastructure (7.4), applied to model choice rather than compute allocation.