Align model selection with task requirements (cost, speed, quality)

Key points

1. 1

   Start from the task, not the tier. Ask: how hard is the reasoning, how costly is an error, how many items, how fast is it needed?

2. 2

   Simple, high-volume, time-boxed work (short summaries, tagging, subject-line rewrites, routing tickets) fits the fastest tier; the most capable tier there wastes time and usage for no visible gain.

3. 3

   One-off, high-stakes, multi-constraint reasoning (an acquisition recommendation, a five-year strategy) fits the most capable tier; speed and volume are irrelevant and the cost of an error dominates.

4. 4

   Mixed everyday workloads fit the balanced tier as a default, stepping up to the most capable tier only for the occasional unusually hard problem.

5. 5

   Fit the model to each step of a process rather than the whole process: fastest for the bulk classification step, balanced for ordinary drafting, most capable for the single strategic analysis.

6. 6

   Consistency of style comes from instructions and examples, not from using one tier everywhere.

7. 7

   Make the decision with evidence when stakes are high: run the candidate tiers on a representative test set, pick the cheapest and fastest one that meets the required accuracy, and re-test when the inputs (such as a handbook) change.

8. 8

   Adding "be extremely careful" to instructions does not close a measured accuracy gap; a stated correctness requirement rules out a tier that misses it, however fast.

9. 9

   Effort and extended thinking add per-response time and usage; they suit hard problems, not bulk simple work under a deadline.

10. 10

    The official sample question for this skill keys the faster, lower-cost model for high-volume short customer replies; always using the top model, disabling features or switching platforms are the distractors.

11. 11

    Common distractors: "volume deserves the top model"; "the top tier never needs checking"; "use the default until it fails" on a costly one-off; alternating tiers item by item.

Read the source

• Change the model, effort, and thinking settings
• Usage limit best practices
• Choosing a model (Anthropic docs)