CCDV-F - Model Selection and Optimization - Section 5.3

Select a Claude model by weighing its capabilities against quality, latency, and cost.

Published skill weight 2.7 percent. Covers the capability profile of Opus, Sonnet and Haiku and the use cases each fits, support for adaptive thinking, the tradeoffs across quality, latency and cost, and the breaking behaviour changes that arrive across model releases and must be accounted for when pinning or upgrading.

Opus, Sonnet, Haiku use casesadaptive thinking supportquality, latency, cost tradeoffsbreaking behaviour changes across releases

Practice question for this objective

Free sampleModel Selection and Optimizationeasy

A pricing team runs a nightly job that reworks 20,000 short product descriptions. The job is judged acceptable on quality today, extended thinking is switched on for every request, and finance has set a cost ceiling per description that the job now exceeds. A separate weekly job that reconciles complex contract clauses depends on the same setting. Which change fits the stated ceiling?

  • ALeave extended thinking on for both jobs and reduce the output token limit on the nightly job until the measured spend per description falls under the stated ceiling.
  • BTurn extended thinking off for both jobs, on the grounds that a setting applied inconsistently between two jobs is harder for the team to reason about and to support.
  • CTurn extended thinking off for the nightly rewriting job, which the team already judges acceptable, and leave it on for the weekly contract reconciliation job. Correct
  • DKeep extended thinking on everywhere and instruct the model in the system prompt to think briefly, so that the nightly job produces fewer reasoning tokens per description.
Extended thinking trades cost for reasoning quality, so enable it per task on difficulty rather than as one setting across every job. Extended thinking lets the model generate reasoning before its answer, and those tokens are produced and paid for on every request that uses it. Applying it to a task already judged acceptable spends that budget for no measured quality gain, while a genuinely hard reconciliation task is where the same spend earns its keep.

Why A is wrong: Tempting because a lower output limit does cap tokens, but it truncates a description mid-generation rather than making the work cheaper, so it buys the ceiling by breaking the deliverable.

Why B is wrong: Tempting because one uniform setting is simpler to operate, but it strips reasoning from the contract reconciliation work, where nothing in the stem suggests quality can be given up.

Why C is correct: Correct because extended thinking buys reasoning quality at the price of extra generated tokens, so it is worth paying only where the task's difficulty needs it, and rewriting short descriptions does not.

Why D is wrong: Tempting because wording can shorten a response, but it leaves the cost lever set by an instruction the model may not hold to, rather than by the switch the team controls.

See more CCDV-F practice questions, answers explained.

More in this domain

Back to all Model Selection and Optimization objectives, or the CCDV-F cert hub.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.