model-tieringlisted

# Model Tiering Use the least capable model tier **and effort** that can succeed at each unit of work. An omitted model flag inherits the parent's model -- usually the most expensive one -- which silently defeats tiering. ## The tier table Tiers are **aliases**, not model IDs. Aliases point to the provider's recommended version and update over time, so the table survives model releases; resolution is provider-dependent (Bedrock, Foundry, and Google Cloud resolve `opus` and `sonnet` differently from the first-party API). A rejected alias exits 1 at launch -- a bad entry is a visible failure, not a silent downgrade. | Work shape | Tier | |---|---| | Transcription plus testing (the plan carries the literal code), or a single-file mechanical fix | `haiku` | | Mechanical execution -- complete spec, 2-3 files, existing pattern to follow | `sonnet` | | Design / triage -- multiple valid approaches, cross-layer, product decisions | `opus` | | Long-horizon autonomous work -- larger than one sitting | `fable` | | Integration -- merge branches, run verification, report | `sonnet` | | Review -- disposable artifact or diff reviewer | `sonnet` | | Simple, high-volume, or disposable lookup | `haiku` | **Cost floor.** Cheapest models take 2-3x the turns on multi-step work and cost more overall. `sonnet` is the floor for reviewers and for prose implementers. `haiku` is only for work where the input already contains the answer: transcription plus testing, single-file mechanical fixes, simp