CCDV-F questions and answers — Domain 2: Model Selection and Optimisation

9 practice questions on Model Selection and Optimisation, the domain weighted at 16.8% of the CCDV-F exam. Every answer explains why the distractors fail — that is the part that transfers to questions you have not seen before.

Question 1

A classification endpoint handles 50,000 requests a day with a simple label set and a strict latency budget. Which tier fits?

1. The highest capability tier, for accuracy
2. A Haiku-class tier, evaluated against your labelled set
3. Whichever tier the rest of the application uses
4. Alternate tiers to balance load

Answer: B. High volume, simple task, latency-sensitive: the lightest tier is the starting point, and the examinable step is confirming it on your own labelled data rather than assuming. Consistency with the rest of the app is not a selection criterion, and alternating tiers makes behaviour unpredictable.

Question 2

Which change reduces cost without changing what the application does?

1. Lowering temperature
2. Adding cache_control to a large, stable system prompt and tool definitions
3. Reducing max_tokens below what the task needs
4. Removing the system prompt

Answer: B. Caching a stable prefix drops those input tokens to 0.1× on subsequent calls and keeps them off your ITPM budget. Temperature does not affect price; truncating output or removing the system prompt changes behaviour, which the question excluded.

Question 3

Which tokens do not count toward your input-tokens-per-minute rate limit?

1. input_tokens
2. cache_read_input_tokens
3. cache_creation_input_tokens
4. Tool definition tokens

Answer: B. Cache reads are excluded from ITPM, so a high cache hit rate raises effective throughput as well as cutting cost. Cache creation tokens do count, as do ordinary input tokens.

Question 4

You need to classify 200,000 archived support tickets by close of business tomorrow. Which surface fits?

1. The standard Messages API with high concurrency
2. The Batches API, matching results on custom_id
3. Streaming requests in a loop
4. The Batches API with stream: true

Answer: B. Bulk, delay-tolerant, overnight: 50% discount, up to 100,000 requests or 256 MB per batch, most completing within an hour and all within 24. Results can return in any order, so match on custom_id. Streaming is not supported inside a batch.

Question 5

Where should a cache_control breakpoint go?

1. On the final user message, which is usually largest
2. On the last content block that stays identical across requests
3. On every block, to maximise coverage
4. On the first block only, since caching is prefix-based

Answer: B. Caching matches on prefix, so the breakpoint marks how far the identical prefix extends. Marking content that changes each request forces a cache write every time. You get four breakpoints, ordered tools → system → messages.

Question 6

Which change invalidates the cached prefix at every level?

1. Adding an image to the latest user message
2. Changing tool_choice
3. Modifying the tool definitions
4. Enabling streaming

Answer: C. Invalidation cascades downward from the level that changed, and tools sit at the top of the hierarchy — so a tool definition change invalidates tools, system and messages. tool_choice and images invalidate messages only; streaming does not affect caching.

Question 7

What is the token-counting endpoint most useful for?

1. Getting the exact billed token count for a completed request
2. Checking a prompt fits the context window and routing small requests to a cheaper model before sending
3. Reducing the number of tokens in a prompt
4. Measuring output tokens

Answer: B. It is a pre-flight estimate over the same inputs a Messages call takes, and it is free. It does not reduce anything by itself, does not measure output, and is not authoritative for billing — usage on the real response is.

Question 8

A team enables a large thinking budget on a high-volume extraction endpoint and costs rise sharply with no quality gain. What is the correct diagnosis?

1. The budget was too small to help
2. Reasoning tokens are billed as output and buy little on extraction, which is not an analytical task
3. Thinking should always be paired with a larger model
4. The budget must exceed max_tokens

Answer: B. Thinking tokens are billed as output and add latency. They pay off on genuinely analytical work and do almost nothing for extraction, formatting or classification. Note also that the budget must sit below max_tokens, with a minimum of 1,024.

Question 9

Your traffic averages well inside your RPM limit but you still get intermittent 429s. Why?

1. Rate limits reset hourly rather than per minute
2. Capacity refills continuously under a token-bucket algorithm, so short bursts can exceed the limit even when the average is fine
3. The limits apply across all models combined
4. 429 always means a spend limit was reached

Answer: B. Token-bucket replenishment means burstiness matters, not just the average. Smooth the traffic, watch the anthropic-ratelimit-*-remaining headers, and honour retry-after. Limits are per model, and a spend-limit 429 is a distinct case that carries no retry-after.

← Previous: Questions & answers — Domain 1  |  Next: Questions & answers — Domain 3 →