Course Content
Domain 1 — Practice Questions
Scenario-based questions on model selection and capability trade-offs
Instructions
These 10 questions mirror the style and difficulty of Domain 1 exam questions. Attempt each question before reading the answer. Track your score — you should target 8/10 or better before moving to the next domain.
Q1. A fraud detection pipeline processes 50,000 transactions per hour and must classify each as suspicious or not within 200ms. Which model is the correct choice?
A) Claude Opus — maximum accuracy for a high-stakes task
B) Claude Sonnet — balanced performance
C) Claude Haiku — lowest latency and cost for high-volume classification
D) A mixture of Opus and Sonnet based on transaction amount
Answer and Explanation
Answer: C — Claude Haiku
This is the canonical Haiku use case: high-volume, latency-sensitive binary classification. 200ms is a hard constraint that only Haiku can reliably meet. Opus would exceed the latency budget and cost ~10x more per transaction. The mixture approach (D) adds complexity without solving the latency problem for Opus calls.
Q2. An analyst needs to compare two 150-page legal contracts (~60K tokens each) and identify all conflicting clauses. This runs once per week. What is the best approach?
A) Use RAG — chunk each contract and retrieve relevant sections
B) Summarize each contract first, then compare the summaries
C) Use Claude Opus with both contracts in context
D) Use Claude Haiku — it is fast enough and weekly frequency reduces cost concerns
Answer and Explanation
Answer: C — Claude Opus with both contracts in context
Both documents total ~120K tokens, well within the 200K context window. Passing both in full preserves all cross-document context that a conflict analysis requires — RAG chunking would destroy this. Opus is correct because identifying conflicting clauses across long legal documents is a high-complexity reasoning task. Weekly frequency means cost is not a primary concern. Haiku (D) lacks the reasoning depth for this task.
Q3. You are building a customer support chatbot that answers questions about your product. 95% of queries are simple FAQs. 5% require nuanced multi-step troubleshooting. Which architecture is most cost-efficient?
A) Use Sonnet for all queries — it handles both cases reliably
B) Use Haiku for all queries — reduce cost across the board
C) Use Haiku by default; escalate to Sonnet when the query is flagged as complex
D) Use Opus for all queries — maximize quality for customer-facing interactions
Answer and Explanation
Answer: C — Haiku by default; escalate to Sonnet for complex queries
This is the model routing pattern. 95% of queries go to Haiku (lowest cost), 5% escalate to Sonnet. Using Sonnet for all (A) wastes cost on simple FAQs. Using Haiku for all (B) risks quality failures on complex queries. Opus (D) is over-engineered for a support chatbot and would not meet latency expectations for the 95% of simple queries.
Q4. A team wants to use extended thinking for their product description generation feature to improve output quality. Is this the right use of extended thinking?
A) Yes — extended thinking always improves output quality
B) No — extended thinking only helps with complex reasoning tasks, not content generation
C) Yes — it will make descriptions more creative
D) No — extended thinking is not available on any Claude model
Answer and Explanation
Answer: B — No, extended thinking only helps with complex reasoning
Extended thinking is designed for problems that require multi-step logical reasoning — architecture analysis, security review, complex deduction. Content generation (product descriptions) does not benefit from extended reasoning chains. Using extended thinking here would add cost and latency with no quality improvement. It is also only available on Opus models, not Sonnet or Haiku.
Q5. Your application currently uses Sonnet and processes 100,000 requests/day with an average of 2,000 input tokens and 1,000 output tokens per request. Switching to Haiku would save approximately what fraction of costs?
A) 10–20%
B) 40–50%
C) 75–90%
D) 99%
Answer and Explanation
Answer: C — 75–90%
Haiku is approximately 8x cheaper on input tokens and 10x cheaper on output tokens compared to Sonnet. At those ratios, switching saves 87.5% on input costs and 90% on output costs — landing in the 75–90% range overall. This is a standard cost estimation question. The exam expects you to know the approximate cost differential between tiers.
Q6. Which of the following tasks is the WORST fit for Claude Haiku?
A) Classifying customer intent from a support ticket
B) Extracting structured fields from a form submission
C) Analyzing a 200-page technical architecture document for security vulnerabilities
D) Routing a user request to one of five internal departments
Answer and Explanation
Answer: C — Security vulnerability analysis of a long technical document
This is a high-complexity reasoning task on a long document. It requires deep reading comprehension, cross-reference analysis, and security domain judgment. Haiku is optimized for simple, high-speed tasks. A, B, and D are all classification or extraction tasks that are correct Haiku use cases.
Q7. A team is debating whether to use extended thinking for a code review feature. The reviewer says: “Extended thinking will make Claude check its reasoning before giving feedback.” Which statement best evaluates this claim?
A) Correct — extended thinking always makes Claude double-check its output
B) Partially correct — extended thinking provides more reasoning tokens, but is best reserved for tasks where reasoning chains are genuinely complex
C) Incorrect — extended thinking only affects response length, not reasoning quality
D) Incorrect — extended thinking is only for math problems
Answer and Explanation
Answer: B — Partially correct
Extended thinking does give Claude more space to reason before responding. However, it is most beneficial for genuinely complex reasoning tasks. For straightforward code review on small files, Sonnet without extended thinking is likely sufficient and cheaper. For reviewing complex security-critical systems or cross-cutting architectural concerns, extended thinking on Opus may be justified. The correct answer recognizes the nuance rather than treating it as a universal improvement or a complete rejection.
Q8. The Anthropic API returns a response with model: "claude-haiku-4-5-20251001". Which of the following is NOT a valid reason to have chosen this model?
A) The task requires classification of user intent
B) The application must respond in under 100ms
C) The task involves comparing and synthesizing three 80-page research papers
D) Cost is the primary constraint and accuracy requirements are moderate
Answer and Explanation
Answer: C — Synthesizing three 80-page research papers
Research synthesis across multiple long documents is a high-complexity task requiring deep reasoning. Haiku would likely produce superficial or inaccurate synthesis. This is a task for Opus (or at minimum Sonnet). All other options (A, B, D) are valid reasons to choose Haiku.
Q9. What is the context window size for all Claude 4.x models?
A) 32K tokens
B) 100K tokens
C) 200K tokens
D) It varies by tier — Haiku has 32K, Sonnet 100K, Opus 200K
Answer and Explanation
Answer: C — 200K tokens for all Claude 4.x models
All models in the Claude 4.x generation share a 200K token context window. Context window size is therefore NOT a differentiator when selecting between tiers. The selection decision is driven by capability, cost, and latency. D is a common distractor — it was true of older generations but not Claude 4.x.
Q10. A company runs nightly batch analysis of 10,000 support tickets (~500 tokens each). Turnaround within 8 hours is acceptable; cost is the top priority. What is the best architecture?
A) Use Opus for maximum analysis quality
B) Use Sonnet for balanced performance
C) Use Haiku via the synchronous API
D) Use Haiku via the Anthropic Batches API for further cost reduction on non-latency-sensitive workloads
Answer and Explanation
Answer: D — Haiku via the Batches API
The Batches API offers additional cost reduction (approximately 50% discount) for non-latency-sensitive workloads. Batch jobs have an up-to-24-hour turnaround, which fits the 8-hour window. Combined with Haiku’s already-low cost, this is the most cost-efficient architecture for a nightly non-interactive workload. Options A and B spend more than necessary; C misses the additional Batches API savings.
Score Interpretation
| Score | Readiness |
|---|---|
| 9–10 / 10 | Domain 1 ready — move to Domain 2 |
| 7–8 / 10 | Review the questions you missed; re-read the model selection framework |
| < 7 / 10 | Re-read the Domain 1 lesson and build the routing pattern hands-on before retesting |