Domain 5: Model Selection and Optimization
Self-check: Model Selection and Optimization
Two separate things. The recall prompts come from this domain's notes and you grade yourself on them; the quiz below is scored against original bank items by the same rule the mock uses. Only the quiz feeds a readiness signal.
Recall prompts
Answer from memory, then grade yourself. Nothing here is scored, and none of it counts towards readiness — the quiz below is the half that does.
Technical Fundamentals
Which single endpoint carries tool use, structured output, caching, thinking, and vision? Name the supporting endpoints that feed or describe it.
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
Name four things an official SDK gives you that a raw HTTP client does not.
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
Which HTTP status codes does the SDK retry by default, and which does it deliberately not retry?
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
At roughly what output size does streaming stop being a preference and become a requirement, and what fails if you ignore that?
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
A workload moves to Amazon Bedrock. What changes besides the credentials?
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
Why does catching a single broad API exception cause an operational problem rather than merely being untidy?
This is measured under Technical Fundamentals. The answer is in that section of this domain's notes above.
LLM Fundamentals
What replaced `budget_tokens` on current models, and what happens if you send it anyway?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
Which model in the current lineup still uses manual extended thinking?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
Where exactly is the effort parameter set? Name the field path.
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
Which effort level is the default, and what is the consequence of setting it explicitly?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
Why does changing effort partway through a conversation cost money beyond the effort change itself?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
You index into `content[0]` expecting a thinking block and it is a text block. What did you assume that adaptive thinking does not guarantee?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
How do you reduce output variability now that `temperature` is removed?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
State two ways fast mode differs from raising effort.
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
A prompt of fixed length is moved from a pre-4.7 model to a current one. What happens to the token count, and what should you do about it?
This is measured under LLM Fundamentals. The answer is in that section of this domain's notes above.
Cost and Token Management
What are the two structural discounts, and what does each cost relative to the base price?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
In what order are `tools`, `system`, and `messages` rendered for caching, and why does the order matter?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
How many cache breakpoints can one request carry?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
`usage.cache_read_input_tokens` is zero across repeated requests with what you believe is an identical prefix. Name three likely causes.
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
Thinking tokens are billed as which kind of token, and which limit do they count against?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
A response stops with `stop_reason: "max_tokens"`. Name the two remedies and the question that decides between them.
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
Does `display: "omitted"` reduce what you are billed? Explain.
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
Which usage field tells you how much of the billed output was reasoning?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
Why is cost per completed task the right unit rather than cost per request?
This is measured under Cost and Token Management. The answer is in that section of this domain's notes above.
Model Selection and Tradeoffs
Name the four current models in capability order, with their per-MTok input and output prices.
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Claude Haiku 4.5 differs from the rest of the lineup on three axes at once. Name all three.
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Which models have a 1M-token context window, and which does not?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
What is the maximum synchronous output for the 1M-context models?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Is `claude-opus-5` an alias or a pinned snapshot? What is the practical consequence?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
List the four breaking changes that make a model migration more than a string swap.
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Your evals fall short on Claude Opus 5 at high effort. What are the two next moves, and in which order would you try them?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Why does a two-model cost cascade forfeit something a single model at lower effort keeps?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
What does a published retirement commitment let you do that an unannounced removal would not?
This is measured under Model Selection and Tradeoffs. The answer is in that section of this domain's notes above.
Cross-cutting
This domain is 16.8% of the exam, roughly 9 items. Which of its four sub-skills carries the most weight, and does your study time reflect that?
This is measured under Cross-cutting. The answer is in that section of this domain's notes above.
For each claim you just recited, could you point to the row in `SOURCES.md` that backs it? An unsourced claim you are confident about is the most dangerous kind.
This is measured under Cross-cutting. The answer is in that section of this domain's notes above.
Scored quiz
This quiz asks 9 items — Model Selection and Optimization's share of a 53-item mock. Domains carry different weights, so quizzes are different lengths.
Scored by the same rule as the mock: an item is right only when your selection is exactly the correct set. This is the only part of this page that feeds a readiness signal.
Result
What went wrong
Every item was answered correctly, so there is nothing to explain.