The current Claude lineup and the model·effort decision palette, in one place. Conceptual charts change rarely; the data strip and curve points refresh monthly from the routine’s research pass — official release tables, pricing pages, and independent benchmarks, never screenshots.
Every recommendation re-derives from these axes. The menu of models and effort tiers is a palette — if recommendations start clustering on one setting, that’s a signal to re-derive, not confirmation.
Capability plus context volume. Heavy context or heavy reasoning pushes up-tier; mechanical work with light context doesn’t need the headroom.
Reasoning budget per turn — thinking depth, tool calls, thoroughness. Difficulty sets its value, independent of how long the work runs.
How likely the conversation surfaces something unforeseen. The tiebreaker that pushes the first two axes up.
The error-economics axis — charted below. How much runs unreviewed, and how many steps compound before anyone sees output.
Supervision lowers the cost of a shallow answer — it never raises its value. A genuinely hard exchange merits high effort even when every turn is reviewed. Weigh this axis with difficulty, never instead of it. These are pressures, not rules.