Choose which Claude model class — Mythos/Fable, Opus, Sonnet, or Haiku — to run a workload on, and use effort level to dial in the quality/speed/cost balance. Use when picking a model for a new production workload, when revisiting a model choice after evals show a gap, when a workload is latency- or cost-sensitive enough that a lower class is worth testing, or when deciding whether to pair a cheaper worker model with a more capable advisor. Default recommendation is to start with the most intelligent generally available model and tune down, because cost-per-task is often lower on more capable models even when price-per-token is higher.
