The core idea: every Claude generation offers a spread of models along the same trade-off triangle — quality, speed, and cost/volume. You cannot maximize all three at once; model selection is choosing which two matter most for a given task.
Durable heuristic (this doesn't change even as specific models do):
- High-volume, simple, latency-sensitive tasks → the fastest, cheapest tier
- General daily professional work → the balanced default tier
- Complex, high-stakes, or ambiguous-reasoning tasks → the top-capability tier(s)
Two things worth knowing that are also durable, not tied to any specific model generation:
- The number of tiers itself has changed over time. Anthropic has run a three-tier lineup in the past; a fourth tier now sits above what used to be the top. Don't assume the tier count is fixed — assume the pattern (fast/cheap → balanced → high-capability) persists even as the exact number and names shift.
- Model IDs are pinned snapshots, not evergreen pointers. A workflow that "just worked" for months can stop behaving the same way after a new model release — not because the old model broke, but because a newer, un-adopted model exists alongside it. Knowing to check "am I still pointed at the model I think I am" is itself a troubleshooting skill.
Comparative framing beats memorized numbers. "The balanced-tier model has a larger context window than the fastest tier" survives version churn. "Model X has exactly Y tokens of context" does not — that's exactly the kind of fact that goes stale. For the exam and for real work, reach for comparative statements first.
→ For the actual current model names, roles, and specs, see Appendix A below.