21 Models, Prompting & Context Engineering Practice Questions & Answers

1. 1. An architect reviews a system where every step — including a high-volume step that only tags incoming messages with one of six labels — is sent to the highest-capability, highest-cost model tier. Quality is acceptable everywhere. What should the architect recommend?

   • A.Test the tagging step on a cheaper, faster tier against a representative evaluation set and move it down if it meets the quality bar✓ Answer
   • B.Leave it as is, because a more capable tier can never produce a worse result than a cheaper one
   • C.Move every step to the cheapest available tier to standardize on one model
   • D.Replace the tagging step with the tier that ranks highest on the latest public leaderboard

   The governing principle is to use the least expensive tier that reliably meets the requirement for that specific task, verified on your own data. Capability purchased beyond what the task needs converts directly into cost and latency with no return, and a narrow closed-set classification is the archetypal case where a cheaper tier usually holds quality. Standardizing on the cheapest tier everywhere is the same error in the opposite direction, because it ignores the steps that genuinely need more capability.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 objective "Select appropriate Claude models based on trade-offs"; Anthropic, "Choosing the right model" (capability / speed / cost / effort criteria)Report a problem with this question

2. 2. A team must choose between two model tiers for a contract-review workload. One engineer proposes deciding based on a widely cited public benchmark ranking. What is the architect's best response?

   • A.Adopt whichever tier the vendor documentation describes with the strongest capability language
   • B.Build an evaluation set of real contract-review cases with known correct outcomes and compare the candidate tiers on it✓ Answer
   • C.Run both tiers in production for a quarter and let user complaints decide
   • D.Adopt the higher-ranked tier, since benchmark ranking is the industry-standard proxy for production quality

   The documented decision procedure puts the evaluation set first: build benchmark tests for your own use case, run the actual prompts and data through each candidate, and compare accuracy, output quality, and edge-case handling before weighing cost. A public leaderboard measures a task distribution that is not yours, so it can rank tiers in an order that reverses on your workload. Waiting for user complaints is not a measurement strategy — it discovers the answer after the cost has already been paid in customer harm.

   Source: Anthropic Claude certification (Architect – Professional), Domain 4 ("optimize token usage, latency, and cost-performance trade-offs"); Anthropic model-selection guidance: "having a good evaluation set is the most important step"Report a problem with this question

3. 3. A cheaper tier scores nearly as well as a more capable tier on published general benchmarks, but performs clearly worse on a company's medical-coding extraction task. What best explains this?

   • A.Extraction tasks are inherently unsuited to any model tier and should be done with rules only
   • B.Published benchmarks aggregate over tasks, prompts, and output formats that differ from this workload, so a near-tie there does not predict parity on a narrow specialized task✓ Answer
   • C.The published benchmark scores must have been measured incorrectly
   • D.The cheaper tier was almost certainly configured with the wrong temperature

   A benchmark score is an average over a task distribution chosen by the benchmark author, and averages hide per-task variance. A workload with unusual vocabulary, a strict output schema, or domain-specific edge cases can sit in exactly the region where two tiers diverge, which is why the selection methodology requires re-measuring candidates on representative in-house data rather than importing an external ranking.

   Source: Anthropic Claude certification (Architect – Professional), Domain 4 — benchmark tests for your use case; Anthropic model-selection guidance on comparing accuracy and edge-case handling on your own dataReport a problem with this question

4. 4. A research team is unsure whether any model can perform a novel multi-step scientific reasoning task at acceptable accuracy. Cost is a secondary concern for the pilot. Which selection strategy fits?

   • A.Skip the pilot and select the tier whose documentation mentions scientific reasoning
   • B.Start with the cheapest tier so the pilot budget stretches to more experiments
   • C.Start with the most capable tier to establish whether the task is achievable at all, then test cheaper tiers against that accuracy ceiling✓ Answer
   • D.Start with a mid-tier model and average its results over many samples to approximate a stronger one

   This is the capability-first strategy, and it exists to separate two different failure causes: a task that no model can do yet, versus a task a cheaper tier simply cannot do. Establishing the ceiling with the strongest tier first means any later quality shortfall is attributable to the downgrade rather than to the task being infeasible, and the optimization to cheaper tiers or lower reasoning effort happens after feasibility is proven.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 — capability-first selection strategy; Anthropic, "Choosing the right model"Report a problem with this question

5. 5. A startup is prototyping a customer-support summarizer. Traffic will be high, per-request latency must stay low, and the summaries are short and formulaic. Which starting strategy is most appropriate?

   • A.Start with the most capable tier so quality complaints never occur, and accept the cost
   • B.Alternate tiers randomly across requests to gather comparative data in production
   • C.Start with a fast, inexpensive tier, measure quality, and upgrade only if a specific capability gap appears✓ Answer
   • D.Start with the most capable tier and never revisit the choice, since prototypes become production

   Efficiency-first selection fits prototyping, latency-bound paths, cost-sensitive economics, and high-volume simple tasks — all four of which describe this workload. Upgrading is then a targeted response to an identified capability gap rather than an untested assumption, which keeps the eventual production configuration justified by evidence.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 — efficiency-first selection strategy; Anthropic, "Choosing the right model"Report a problem with this question

6. 6. Analysis shows that roughly 85% of incoming requests are routine and 15% are genuinely hard. The team wants to cut cost without losing accuracy on the hard cases. Which design best serves this, and what does it require?

   • A.Send all traffic to the more capable tier, since a routing layer always costs more than it saves
   • B.Send all traffic to the cheaper tier and accept degraded results on the hard 15%
   • C.Send routine traffic to a cheaper tier and escalate to a more capable tier on low confidence or a failed validation check, while evaluating the routing decision itself end to end✓ Answer
   • D.Let each end user choose which tier handles their request

   A cascade captures most of the savings because the cheap tier absorbs the majority of traffic while the escalation path preserves accuracy where it matters. The cost of the pattern is real, though: the router or confidence check becomes a component that can misroute, so the correct answer includes evaluating the combined system's accuracy and total cost rather than only the cheap path's unit price.

   Source: Anthropic Claude certification (Architect – Professional), Domain 1 — heterogeneous multi-tier architectures and escalation cascades; Domain 4 — cost-performance trade-offsReport a problem with this question

7. 7. An agent must run autonomously for hours across dozens of dependent steps, where an early mistake silently corrupts everything downstream. A cheaper tier passes 95% of single-step tests. What does that 95% imply for tier selection?

   • A.The chain length has no bearing on tier selection; only the difficulty of the hardest single step matters
   • B.Single-step accuracy is irrelevant to agents and should not be measured
   • C.A 95% single-step pass rate is sufficient, since 95% exceeds most quality targets
   • D.Per-step error compounds across a long dependent chain, so a long-horizon autonomous task justifies a more capable tier even at higher unit cost✓ Answer

   Task horizon is one of the primary selection questions precisely because independent per-step errors multiply: a 95% per-step success rate over thirty dependent steps yields a task-level success rate far below 95%. Long-running, high-autonomy work with unrecoverable downstream corruption is therefore the canonical case for spending more per token, since the cost of a failed run dwarfs the token savings.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 — capability-first selection for high-autonomy agentic work; Domain 1 — agent viability criteria (errors must be recoverable)Report a problem with this question

8. 8. A document pipeline has two stages: pull named fields out of a fixed form layout, then decide whether the extracted terms breach a policy. How should tiers be assigned?

   • A.Use the cheapest tier for both and add a retry loop on each stage
   • B.Use one tier for both stages so the pipeline stays simple to operate
   • C.Use a cheaper tier for the structured extraction and a more capable tier for the policy judgment✓ Answer
   • D.Use a more capable tier for extraction, because extraction errors are the root cause of everything downstream

   Model choice is made per stage, not per pipeline, because the stages have different difficulty profiles: reading known fields from a fixed layout is a mechanical task a cheaper tier handles reliably, whereas policy judgment is open-ended interpretation where capability actually changes the answer. Assigning tiers stage by stage captures most of the cost saving while spending capability only where reasoning quality is the binding constraint.

   Source: Anthropic Claude certification (Architect – Professional), Domain 1 — decomposition and heterogeneous multi-tier design; Domain 2 — model trade-off selectionReport a problem with this question

9. 9. A workload's outputs are correct about 90% of the time on its current tier; the target is 96%. The prompt is already well structured. Which lever should the architect try before changing tiers?

   • A.Send each request three times and take the majority answer as the permanent design
   • B.Increase the maximum output length so the model has room to be more accurate
   • C.Immediately move to the highest-capability tier available
   • D.Raise the reasoning/effort setting on the current tier and re-measure on the evaluation set✓ Answer

   The reasoning/effort control is a within-model lever that trades additional computation for output quality, and the documented guidance is that tuning it is often a better first move than switching models. It preserves everything already validated about the current configuration, so if the gap closes, the team avoids a migration and a full re-evaluation; only if effort tuning plateaus below the target does a tier change become the justified next step.

   Source: Anthropic Claude certification (Architect – Professional), Domain 2 — effort as a selection criterion; Anthropic, "Choosing the right model": tuning effort is often a better lever than switching modelsReport a problem with this question

10. 10. An architect raises the reasoning/effort setting to its maximum for a straightforward data-normalization task. Accuracy is unchanged, but latency and token usage rise sharply. What is the correct conclusion?

    • A.The evaluation set must be flawed, since more computation always improves accuracy
    • B.More reasoning effort is not universally better; the right level is found empirically per task, and this task should run at a lower setting✓ Answer
    • C.The result proves the tier is too weak and the workload should move up a tier
    • D.The maximum setting should be kept anyway, because it protects against future edge cases

    Effort buys extra deliberation, which only helps when the task actually contains something to deliberate about; a deterministic normalization has no ambiguity for the extra computation to resolve, so the spend converts entirely into latency and tokens. The methodology is to sweep the setting against your evaluation set and pick the lowest level that holds quality, rather than assuming a monotonic relationship between compute and accuracy.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2 — effort trade-offs; Domain 4 — latency and cost-performance optimizationReport a problem with this question

11. 11. A telemetry service must classify roughly 40,000 events per hour into a fixed set of severity levels, with a sub-second per-event budget. Accuracy on a labeled sample is equal across the candidate tiers. Which choice is architecturally correct?

    • A.The most capable tier at a reduced reasoning effort, to balance the two concerns
    • B.The most capable tier, because severity classification affects incident response
    • C.Whichever tier the team already uses elsewhere, to minimize operational variety
    • D.The faster, cheaper tier, because when quality is equal the remaining criteria are speed and cost✓ Answer

    The four selection criteria are capability, speed, cost, and effort, and they are applied in that order of elimination: once measured capability is tied, the decision falls entirely to the remaining three. High-volume, latency-bound classification into a closed label set is the archetypal fit for a fast inexpensive tier, and paying for unused capability at 40,000 events per hour multiplies a per-event premium into a large recurring bill.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2 — capability, speed, cost, effort criteria; Anthropic, "Choosing the right model"Report a problem with this question

12. 12. After downgrading a workload to a cheaper tier, per-token spend drops 60% but the monthly bill falls only 8%. Logs show many requests now run two or three times before passing validation. What metric should have driven the decision?

    • A.Price per million tokens, applied to the projected token volume
    • B.The published context window and maximum output size of each tier
    • C.The number of distinct model tiers the platform makes available
    • D.Total cost per successfully completed task, including retries, escalations, and human review✓ Answer

    Unit token price is an input cost, not an outcome cost; the economically meaningful quantity is what it costs to get one correct finished result. A cheaper tier that needs more attempts, longer prompts, extra pipeline steps, or human correction can be more expensive per completed task than the tier it replaced, which is why cost comparisons must be run end to end on the evaluation set before a downgrade is accepted.

    Source: Anthropic Claude certification (Architect – Professional), Domain 4 — "optimize token usage, latency, and cost-performance trade-offs"; cost-per-completed-task measurementReport a problem with this question

13. 13. A legal analysis task requires reasoning over a corpus far larger than any candidate model's context window. What is the sound design response?

    • A.Treat the context window as a fixed budget and design selection, summarization, or staged processing so only relevant material occupies it✓ Answer
    • B.Truncate the corpus arbitrarily until it fits and accept whatever is lost
    • C.Choose the tier with the largest advertised context window and send the entire corpus in one request
    • D.Wait for a future release with a larger window before building the system

    A context window is a hard, finite budget, so a task that exceeds it is a design problem rather than a shopping problem: the architecture must decide what earns a place in the window. Selection, summarization, and staged processing preserve the material that matters, whereas arbitrary truncation silently discards evidence the answer may depend on and picking a bigger window merely postpones the same limit.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2 — context engineering; Domain 1 — decomposition of tasks that exceed a single requestReport a problem with this question

14. 14. An architect is deciding between one long request that handles an entire report end to end and a decomposed pipeline of smaller steps. Which statement best captures the trade-off?

    • A.The choice is purely a matter of developer preference and has no measurable consequences
    • B.Decomposition is always superior because smaller requests are cheaper
    • C.Decomposition adds checkpoints, retry granularity, parallelism, and per-stage tier choice, at the cost of orchestration complexity and some loss of cross-cutting context✓ Answer
    • D.A single long request is always superior because it preserves full context

    Decomposition converts one all-or-nothing call into stages that can each be retried, parallelized, validated, and assigned its own tier, which is what makes mixed-tier cost optimization possible at all. The price is real orchestration overhead and the risk that a stage loses context another stage needed, so the decision is made per workload by weighing recoverability and cost against the value of holding everything in one reasoning pass.

    Source: Anthropic Claude certification (Architect – Professional), Domain 1 — apply decomposition; Domain 4 — cost and latency optimizationReport a problem with this question

15. 15. Users describe a chat assistant as 'sluggish,' though total response time meets the internal target. Responses are returned only after the full answer is generated. What should the architect address first?

    • A.Stream the response so output appears as it is produced, which reduces perceived latency without changing total generation time✓ Answer
    • B.Cap the answer length so every response finishes sooner regardless of the question
    • C.Switch to a faster, cheaper tier and accept lower answer quality
    • D.Tell stakeholders the target is met and close the issue

    Latency must be evaluated as the user experiences it, and for interactive interfaces the dominant felt quantity is time to first visible token, not total completion time. Streaming changes the experience without changing the work performed, so it should be exhausted before trading away quality by moving to a different tier — a downgrade would address a metric that is already within target.

    Source: Anthropic Claude certification (Architect – Professional), Domain 4 — latency optimization and user-perceived latency; Domain 6 — SLA alignmentReport a problem with this question

16. 16. A regulated workflow must be able to reproduce, months later, the exact configuration that generated an archived decision. What is the correct model-versioning practice?

    • A.Always call the latest version so archived decisions reflect the best available reasoning
    • B.Store the output text alone; the model version is not part of the audit record
    • C.Record only the model family name, since versions within a family behave identically
    • D.Pin an explicit model version in configuration and record it, along with the settings used, on every archived decision✓ Answer

    Reproducibility requires that every input to the result be captured, and the model version and its generation settings are inputs. A floating alias silently changes behavior underneath an unchanged application, so a decision archived under 'latest' cannot be reconstructed; pinning plus recording makes the configuration an auditable fact rather than an assumption.

    Source: Anthropic Claude certification (Architect – Professional), Domain 6 — documenting architectures and lifecycle management; model version pinning for reproducibilityReport a problem with this question

17. 17. A vendor announces a newer version within the same model family, described as an improvement over the version currently in production. What should the team do before switching?

    • A.Switch immediately but keep the old prompts unchanged, since prompts are model independent
    • B.Switch immediately, because a newer version in the same family is by definition backward compatible in behavior
    • C.Skip the upgrade entirely; migrations are avoidable if the current version still works
    • D.Re-run the existing evaluation suite on the new version, compare results including edge cases, and roll out gradually before retiring the old version✓ Answer

    A version change is a behavior change even when average quality improves, because gains on aggregate metrics can coexist with regressions on specific edge cases, output formats, or instruction-following habits your prompts depend on. The prescribed procedure is to test the new version against your own evaluation set with your real prompts and data, compare accuracy and edge-case handling, and stage the rollout so a regression is caught before it reaches all traffic.

    Source: Anthropic Claude certification (Architect – Professional), Domain 4 — build benchmark tests and compare across models before changing; model migration guidanceReport a problem with this question

18. 18. How should an architect treat model deprecation and retirement in a long-lived production system?

    • A.As a vendor problem that does not affect architecture decisions
    • B.As a reason to avoid model-based components in long-lived systems altogether
    • C.As an emergency to handle when requests begin failing in production
    • D.As a normal lifecycle event: track announced dates, keep the evaluation suite maintained, and schedule migration windows in advance✓ Answer

    Models are announced, deprecated, and retired on published schedules, so continued availability is a time-limited assumption rather than a permanent property of the design. Treating migration as planned work — with a maintained evaluation suite ready to re-run and a scheduled window — turns a forced cutover into a routine, tested change instead of an incident discovered through failing requests.

    Source: Anthropic Claude certification (Architect – Professional), Domain 6 — lifecycle management (monitoring, iteration); model deprecation and retirement policyReport a problem with this question

19. 19. During design review, an engineer asserts a specific capability limit for a model tier, citing a comparison article read some months ago. What should the architect do?

    • A.Accept the figure, since capability limits rarely change once published
    • B.Ask a different engineer whether they remember the same figure
    • C.Verify the limit against the current official model documentation or a live capability query before the design depends on it✓ Answer
    • D.Design for half the stated limit as a safety margin without checking

    Capability figures move with each release, and third-party comparison articles routinely describe a lineup that has since changed, so a remembered number is an unverified input to a design. The methodology is to treat published capabilities as something to look up at design time — from current official documentation or a programmatic capability query — rather than something to memorize, which also keeps the design from silently encoding a stale limit.

    Source: Anthropic Claude certification (Architect – Professional), Domain 7 — live capability discovery instead of hard-coded tables; Domain 6 — documenting and validating design assumptionsReport a problem with this question

20. 20. A cheaper tier meets the quality bar on 97% of cases in an eligibility-decision workflow, but the remaining 3% produce denials that are costly to reverse and damaging to the applicant. How should this affect tier selection?

    • A.Tier selection should ignore error consequences, which belong to the operations team
    • B.The cost of an individual error is a primary selection input, so the higher error cost justifies a more capable tier and a review step for the decisions at risk✓ Answer
    • C.The 97% pass rate settles it; the cheaper tier should be adopted for the whole workflow
    • D.The workflow should be run on both tiers permanently and disagreements discarded

    Selection is driven by the expected cost of being wrong, not by the pass rate alone: a 3% error rate is acceptable when errors are cheap and self-correcting, and unacceptable when each one is expensive and hard to reverse. Because the consequence is asymmetric here, the justified answer combines a more capable tier with human review on the at-risk decisions, since accuracy alone never fully eliminates the residual error.

    Source: Anthropic Claude certification (Architect – Professional), Domain 2 — accuracy-over-cost selection when errors are costly; Domain 5 — human-in-the-loop validationReport a problem with this question

21. 21. An executive asks why the team chose a mid-cost tier instead of the most capable one. What is the strongest way for the architect to justify and document the decision?

    • A.Present evaluation results showing the chosen tier meets the quality and latency targets at a lower unit cost, and define the evidence that would trigger re-evaluation✓ Answer
    • B.Commit to upgrading to the most capable tier if anyone ever complains about quality
    • C.State that the chosen tier is the industry standard for this category of application
    • D.Explain that the most capable tier was unavailable at the time of the decision

    A model choice is defensible when it is tied to measured performance against the stated targets and to an explicit re-evaluation trigger, which turns it from an opinion into a reviewable engineering decision. Appeals to industry norms or promises to escalate on complaint provide no evidence that the target is met and leave the team without a defined condition for revisiting the choice as workloads and the model lineup change.

    Source: Anthropic Claude certification (Architect – Professional), Domain 6 — communicating architectural decisions and trade-offs; Domain 4 — evidence-based cost-performance justificationReport a problem with this question

Practice questions based on the official Claude Certified Architect – Professional (CCAR-P) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. The real exam is 63 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($175). Official certification page →