22 Models, Prompting & Context Engineering Practice Questions & Answers

1. 1. A team compares two configurations for an agentic ticket-resolution workload. The cheaper-per-token setup resolves 77% of tickets and takes more turns and retries; the pricier-per-token setup resolves 88% in fewer turns. Finance asks which one is cheaper to run. How should the architect frame the comparison?

   • A.Compare the published input and output token rates, because per-token price does not vary from run to run.
   • B.Compare the total spend of each setup divided by the tickets it actually resolved✓ Answer
   • C.Compare the two pilots' monthly invoices directly, since the pilot that billed less is the setup finance should fund.
   • D.Compare the average tokens each setup consumes per request, since a smaller request footprint is what scales cheaply.

   The governing unit for a model decision is cost per completed task, not cost per token. A model with a higher per-token price that finishes the job in fewer turns, with fewer retries and fewer abandoned runs, can produce a lower total cost per resolved ticket; per-token rates and per-request token counts say nothing about how many attempts a task takes. Raw invoices are also uninformative unless the two pilots handled the same volume of successfully completed work.

   Source: CCAR-P Exam Guide v1.0, Domain 2 (Claude Models, Prompting & Context Engineering) — select models based on trade-offs; Anthropic, 'Choosing the right model' / 'Optimizing for cost and intelligence' — cost per completed taskReport a problem with this question

2. 2. A support-summarization service sends a large, unchanging policy corpus with every request and has never enabled prompt caching. The bill has doubled. The team's first proposal is to move the route to a smaller, cheaper model. What should the architect do first?

   • A.Move the route to the smaller model now and run the quality eval afterwards to see whether the summaries still hold up.
   • B.Turn on prompt caching for the stable corpus and re-measure, since that lever cuts spend without costing quality.✓ Answer
   • C.Add per-request cost logging and a spend alert so the team can watch the route before making any configuration change.
   • D.Lower the reasoning effort on the route first, since effort is the cheapest knob and needs no change to the payload.

   Cost levers have an order: the free wins that cost no quality come before any lever that trades capability. A stable prefix repeated on every request is exactly the case prompt caching exists for, and a cache read costs materially less than an uncached read of the same tokens. Lowering effort and switching models both trade quality and should be measured only after the free win is in place; logging observes the problem without reducing it.

   Source: CCAR-P Exam Guide v1.0, Domain 4 (Evaluation, Testing & Optimization) — optimize token usage, latency and cost-performance trade-offs; Anthropic, 'Optimizing for cost and intelligence' — lever order (free wins before trade-offs)Report a problem with this question

3. 3. A document-review pipeline runs at the default effort on the most capable model and costs more than budgeted. An engineer proposes a two-model cascade in which a cheaper model handles easy documents and escalates hard ones. What is the better first experiment?

   • A.Add a monitoring hook that records token spend per document so the team can see which documents drive the overage.
   • B.Build the cascade behind a flag and compare it against the current pipeline on a full week of live traffic.
   • C.Run the same model at a lower reasoning effort against the existing eval set and see whether accuracy holds.✓ Answer
   • D.Raise the request timeout and retry ceiling so hard documents finish on the first attempt instead of a second run.

   Reasoning effort is the cheaper experiment because it changes one parameter on an architecture that already exists, and lower effort on a strong model frequently holds accuracy while cutting spend. A cascade adds a routing rule to maintain, a second failure mode when routing is wrong, and a second cache namespace, since cache entries are scoped to the model that wrote them. Timeouts and retry ceilings address neither cost nor accuracy here.

   Source: CCAR-P Exam Guide v1.0, Domain 2 — model selection trade-offs and tuning effort; Anthropic, 'Optimizing for cost and intelligence' — measure lower effort on the strong model before building a multi-model pipelineReport a problem with this question

4. 4. An assistant sends a large stable system prompt followed by the user's turn. Cache-read token counts are zero on every request even though the prompt text appears unchanged. Inspection shows the system prompt opens with a line rendering the current date and time. What is the fix?

   • A.Round the rendered timestamp to the nearest hour so requests made within the same hour produce identical text.
   • B.Move the rendered timestamp out of the system prompt and into the user turn that follows the last breakpoint.✓ Answer
   • C.Switch the cache entry to the longer time-to-live option so it survives between requests despite the changing line.
   • D.Add a cache breakpoint immediately after the timestamp line so the rest of the system prompt is cached on its own.

   Prompt caching is a prefix match, so any byte that changes invalidates the entry from that point onward. A timestamp at the very top of the system prompt changes the first bytes of the prefix, so nothing after it can ever be reused. Volatile content belongs after the last breakpoint. A longer time-to-live keeps an entry alive but cannot make a changed prefix match, and a breakpoint placed after the volatile line still sits behind bytes that differ on every request.

   Source: CCAR-P Exam Guide v1.0, Domain 2 — optimize context windows and token usage; Anthropic prompt caching documentation — prefix match, silent invalidators, placement of volatile contentReport a problem with this question

5. 5. A team wants a durable caching layout for a multi-turn agent that carries a fixed tool set, a long system prompt, a retrieved case file that changes per conversation but not within one, and the user's question. How should these be arranged?

   • A.Put the user's question ahead of the case file so the model reads the request before the reference material it will need.
   • B.Lead with the case file, ahead of the tool set and prompt, so the largest block is cached before the smaller ones.
   • C.Place a breakpoint after each of the four blocks so every segment is cached and invalidated on its own schedule instead.
   • D.Keep the tool set and system prompt first, put the per-conversation case file after them, and set the last breakpoint after it.✓ Answer

   Content renders in the order tools, then system, then messages, and the cache matches on prefix, so blocks must be ordered from most stable to most volatile. The tool set and system prompt are stable across every conversation, the case file is stable within one conversation, and the question changes every turn, which is why the last breakpoint belongs after the case file. Per-block breakpoints do not give segments independent lifetimes, because a change in any earlier block invalidates everything after it.

   Source: CCAR-P Exam Guide v1.0, Domain 2 — prompt reuse and context optimization; Anthropic prompt caching documentation — render order (tools → system → messages) and breakpoint placementReport a problem with this question

6. 6. A conversational product serves the first turns of a session with a small model and escalates the same conversation to a larger model when the question gets complex. The team expected the escalated turn to read from the cache built during the earlier turns. Why does that not happen?

   • A.Cache entries expire as soon as tool results are returned, so a conversation that calls tools keeps no reusable prefix.
   • B.Cache entries are scoped to the model that wrote them, so the escalated turn writes a new entry instead of reading one.✓ Answer
   • C.Cache entries are invalidated whenever the assistant appends a turn, so only the first request of a session can read one.
   • D.Cache entries are scoped to a single request, so every later request rebuilds the prefix no matter which model serves it.

   Caches are model-scoped: an entry written while serving one model is not readable by another, so switching models mid-conversation forfeits every accumulated cache read and pays a fresh write on the new model. That is a real and often overlooked cost of cascades and routers, and it should be weighed against their projected savings. Appending assistant turns and returning tool results both extend the prefix rather than destroying it, so neither ends reuse on its own.

   Source: CCAR-P Exam Guide v1.0, Domain 2 — model selection trade-offs; Anthropic prompt caching documentation — cache entries are scoped per modelReport a problem with this question

7. 7. After adding cache breakpoints to a route, an engineer reports the change made things worse: the first request after deployment showed a higher input cost than before. How should the architect respond?

   • A.Revert the breakpoints, since a configuration that raises input cost on its first request will not recover the difference.
   • B.Review the request-building code, since a caching configuration that is written correctly never raises input cost at all.
   • C.Compare usage figures from the second request onward, because the first request pays the write and only later ones read the entry.✓ Answer
   • D.Move the breakpoints later in the prompt, since a write that expensive means far too much content was marked cacheable.

   A cache write costs somewhat more than an uncached read of the same tokens, while a cache read costs materially less, so the first request after deployment is expected to look like a regression and the saving only appears from the second request onward. Caching must be verified from the usage fields that report cache creation and cache read tokens, measured across repeated requests rather than judged from a single call or from reading the code.

   Source: CCAR-P Exam Guide v1.0, Domain 4 — monitor with logging and observability; Anthropic prompt caching documentation — cache write vs. cache read pricing direction and verification via usage fieldsReport a problem with this question

8. 8. A nightly job re-classifies the previous day's support tickets and an analyst reads the results the next morning. A second workload classifies a ticket while the customer waits on the page. Which use of batch processing is appropriate?

   • A.Send the nightly re-classification through batch processing and leave the customer-facing path on standard requests.✓ Answer
   • B.Send the customer-facing path through batch and the nightly job on standard requests, since the nightly job is larger.
   • C.Send both workloads through batch processing and poll often enough that the customer-facing path still answers quickly.
   • D.Keep both workloads on standard requests, because batch results come back in an order the caller cannot control.

   Batch processing trades latency for cost: the work is queued and returned asynchronously, so it fits any workload where no user is waiting and fails any workload where one is. Polling more often does not shorten the queue. The out-of-order return of results is a real property, but it is handled by keying results on the caller's own identifier rather than by position, so it is not a reason to avoid batch for the nightly job.

   Source: CCAR-P Exam Guide v1.0, Domain 4 — cost-performance trade-offs; Anthropic Message Batches documentation — asynchronous processing, results keyed by the caller's custom identifierReport a problem with this question

9. 9. A team migrating an older service asks how to carry over its fixed per-request thinking-token budget, which they hand-tuned for their hardest cases. What should the architect advise?

   • A.Keep the fixed budget as a ceiling and add an effort setting above it, so the route keeps a hard cap and gains the newer control.
   • B.Reproduce the budget by setting the response token cap to the same number, since that field enforces the same ceiling on depth.
   • C.Drop the budget and leave reasoning depth unspecified, since any explicit control over depth costs quality on the hardest cases.
   • D.Replace the fixed budget with adaptive reasoning and an effort setting, since the budget is the older form of that control.✓ Answer

   Adaptive reasoning is the modern expression of the same idea: instead of a caller-chosen fixed allowance applied uniformly, the model varies depth per request and the architect sets an effort level for the route. A hand-tuned budget sized for the hardest cases overspends on the easy majority. The response token cap is a separate, enforced ceiling on the response that the model is not shown, so it cannot substitute for a reasoning-depth control.

   Source: CCAR-P Exam Guide v1.0, Domain 2 — model configuration trade-offs; Anthropic extended thinking documentation — adaptive thinking supersedes a fixed thinking-token budget; effort as the depth controlReport a problem with this question

10. 10. An organization runs three Claude routes: an intent classifier on every inbound message, a chat assistant, and a long-horizon code-migration agent. A platform engineer proposes setting one effort level in shared configuration for all three. What is the better design?

    • A.Set effort per route: the low setting on the high-volume classifier, the higher one on the long-horizon agent.✓ Answer
    • B.Set one high effort level for all three, since the classifier's short outputs make the extra reasoning cheap.
    • C.Set one low effort level for all three and raise the response cap on the agent so it still has room to finish.
    • D.Leave effort unset everywhere and cap spend with per-route rate limits, so no route exceeds its share of budget.

    Which workloads repay higher effort is a property of the workload, not of the organization: long-horizon and agentic work gains materially from deeper reasoning, while classification and other high-volume or latency-sensitive routes usually gain little and pay for the extra tokens on every call. Effort therefore belongs in per-route configuration. Rate limits cap exposure but do not change the cost or quality of any individual request.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — optimize cost-performance trade-offs; Anthropic, 'Optimizing for cost and intelligence' — effort by workload type, tuned per routeReport a problem with this question

11. 11. An architect wants to lower the default effort on a customer-facing extraction route. The vendor documentation reports that the lower setting costs materially less with a small accuracy drop on published benchmarks. What is the defensible basis for making the change?

    • A.The published benchmark deltas, since they were measured on a broader task set than any one team could assemble.
    • B.A side-by-side comparison on the three requests the team found hardest, since the hard cases are what decide the question.
    • C.The absence of customer complaints during a week of quietly running the lower setting alongside the current one.
    • D.A run of both settings over a sample of the route's own recent production requests, scored against its eval set.✓ Answer

    A model or configuration decision is defended with a measurement on the actual workload, never with a general claim about which setting is better. Published benchmark deltas describe a different task mix and do not predict this route's behaviour. Three hand-picked requests are too few to attribute a difference, and absence of complaints is a lagging, insensitive signal that misses silent extraction errors nobody happened to notice.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — design evaluation datasets, conduct A/B testing and iterative improvement; Anthropic, 'Optimizing for cost and intelligence' — measure on real traffic before changing a defaultReport a problem with this question

12. 12. A team proposes a router that inspects each incoming request and sends the easy ones to a cheaper model. The workload's difficulty is not evident from the request text alone. What is the strongest argument against building the router now?

    • A.A cheaper model cannot handle any part of a workload whose hardest cases need the larger one, so the split saves nothing.
    • B.The rule has to be maintained and re-validated as traffic and models change, and every misrouted hard case pays for a second run.✓ Answer
    • C.A router cannot be evaluated until it is built, so the team commits to an architecture whose accuracy nobody can estimate.
    • D.Routing to different models within one conversation is unsupported, so any escalation would have to start the session over.

    A router is a component with its own accuracy, its own maintenance burden and its own failure mode, and it only pays when the executor can actually tell hard cases from easy ones cheaply. Here it cannot, so misrouted requests are paid for twice and the rule needs re-validation on every traffic or model change. A router can be evaluated before shipping by replaying historical requests, and switching models across turns is possible; the cost is the forfeited cache, not an outright prohibition.

    Source: CCAR-P Exam Guide v1.0, Domain 1 (Solution Design & Architecture) — select the simplest architectural tier that meets the need; Anthropic, 'Optimizing for cost and intelligence' — when a multi-model split does not payReport a problem with this question

13. 13. A stakeholder asserts that a smaller model would be obviously good enough for the FAQ-answering route. How should the architect settle the question?

    • A.Have the larger model grade a sample of the smaller model's answers, and adopt it if most of those grades come back good.
    • B.Compare the two models' published capability tiers and adopt the smaller one, since answering an FAQ is an easy task.
    • C.Adopt the smaller model on the route and watch thumbs-down feedback, reverting if the negative rate climbs above today's.
    • D.Score both models on a held-out set of real FAQ requests with graded answers, and adopt the smaller one only if it clears the bar.✓ Answer

    Sufficiency is demonstrated, not assumed: the smaller model is adopted only after it clears a stated quality bar on a held-out set of the route's own requests with known-good answers. Published capability tiers describe general standing and not this task. Shipping first and watching feedback puts customers inside the experiment and detects only complaints, and grading by a larger model with no ground truth measures agreement rather than correctness.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — define evaluation metrics and design evaluation datasets; Anthropic, 'Choosing the right model' — validate a downgrade against your own evaluation setReport a problem with this question

14. 14. A service hardcodes its model identifier at six call sites and its prompts were tuned two model generations ago. The team wants to be ready for the next model release. What should they change?

    • A.Pin the identifier at each call site and freeze it, so no release can change behaviour until a migration is scheduled.
    • B.Wrap each call site in a retry that falls back to the previous identifier whenever a call to the new model fails.
    • C.Move the identifier into configuration and keep an eval set that can be run against a candidate before and after.✓ Answer
    • D.Read the identifier from the model listing at startup and always take the newest entry, so releases are tracked for free.

    Model choice should live in configuration rather than being an assumption baked into call sites, and a switch should be gated by the same eval run before and after so the change is attributable. Prompts tuned for an older generation frequently need revisiting, because instructions that helped one model can constrain a newer one. Auto-adopting the newest listing entry changes behaviour with no evaluation, and a retry fallback masks a quality change rather than measuring it.

    Source: CCAR-P Exam Guide v1.0, Domain 6 (Stakeholder Communication & Lifecycle Management) — support lifecycle phases (design, handoff, monitoring, iteration); Anthropic model migration guidance — re-evaluate prompts tuned for an earlier modelReport a problem with this question

15. 15. During a design review, an architect is asked to state the context window and per-token price of each candidate model from memory to fill in a comparison table. What is the right response?

    • A.Leave the table out and compare the candidates only by the results they produce on the team's own eval set.
    • B.Fill the table from the current model listing and pricing documentation, and record the date the figures were read.✓ Answer
    • C.Fill the table from memory and mark it provisional, since the relative ordering of the tiers is what the decision turns on.
    • D.Fill the table from last quarter's design document, since those figures were verified at the time of that review.

    Capability, latency, price and context capacity are the axes that differ between models, but their current values are discovered rather than remembered: the model listing endpoint and the current pricing documentation are authoritative, and a released model can change the numbers at any time. Recording the read date tells a later reader how stale the table is. Comparing only on eval results ignores price and context limits, which are genuine design constraints.

    Source: CCAR-P Exam Guide v1.0, Domain 6 — document architectures and provide implementation guidance; Anthropic Models API documentation — query the model listing for current context and capability valuesReport a problem with this question

16. 16. A healthcare customer requires inference to run in a specific geography and requires that the deployment platform come from their approved list. The most capable candidate model is not offered on that platform in that geography. How should the architect proceed?

    • A.Deploy the most capable model behind a system-prompt rule forbidding storage or transmission of patient information.
    • B.Choose the most capable model and open an exception request with the customer's compliance team while the build proceeds.
    • C.Adopt the most capable model and add response logging with a review queue so any issue is caught before a patient sees it.
    • D.Filter the candidates to models offered on the approved platform in that geography, then compare the survivors on cost and quality.✓ Answer

    Compliance constraints filter the candidate set before any cost or capability comparison begins, because inference geography and platform availability are structural facts about where the workload may legally run, not risks to be mitigated after the fact. Features and models available on one platform may simply not exist on another. A system-prompt rule is an instruction the model can fail to honour, and logging with review detects a violation only after it has already occurred.

    Source: CCAR-P Exam Guide v1.0, Domain 5 (Governance, Safety & Risk Management) — ensure compliance with regulations (GDPR, HIPAA, FedRAMP); Anthropic platform availability documentation — model and feature availability varies by platform and inference geographyReport a problem with this question

17. 17. A checkout-page assistant has a contractual p95 latency target. The configuration that scores highest on the quality eval misses that target on a third of requests. What should the architect bring to the stakeholder review?

    • A.A comparison of the tested configurations on quality and latency, with a recommendation that meets the target.✓ Answer
    • B.An offer to ship the highest-scoring configuration alongside an alert that fires whenever the target is breached.
    • C.A proposal to ship the highest-scoring configuration with a prompt instruction telling the model to answer briefly.
    • D.A request to reclassify the latency target as an internal goal so the highest-scoring configuration can ship as it is.

    A contractual latency target is a hard design constraint, so it legitimately rules out the most capable configuration, and the defensible recommendation is the one that meets the constraint with its quality cost stated in business terms and tied to a measurement. An alert reports breaches after customers experience them, a brevity instruction is an unreliable latency control, and quietly downgrading a commitment is a renegotiation the stakeholders have not agreed to.

    Source: CCAR-P Exam Guide v1.0, Domain 6 — communicate architectural decisions and trade-offs; manage expectation alignment including SLAsReport a problem with this question

18. 18. A workload extracts fields from one contract, validates them against one policy document, then drafts a single summary. Every step depends on the previous one and all the material fits in one context window. An engineer proposes an orchestrator with three subagents. What is the better design?

    • A.Retain the orchestrator and add a shared memory store, so the three subagents avoid re-reading the same contract independently.
    • B.Collapse the design into one sequential workflow in a single context, since no step can start before the previous finishes.✓ Answer
    • C.Reduce the orchestrator to two subagents by merging extraction with validation, so the final merge has fewer results to reconcile.
    • D.Keep the three subagents but give each of them the cheapest model, so the added planning and merge turns cost as little as possible.

    Multi-agent orchestration pays only when there is genuinely parallel bulk work or work that exceeds a single context window. On a strictly dependent chain that fits in one context, the orchestrator pays for planning, dispatch and merge turns that a single sequential run gets for free, and it adds coordination failure modes. Cheaper subagent models, fewer subagents and a shared store all reduce the overhead of a structure that should not exist here.

    Source: CCAR-P Exam Guide v1.0, Domain 1 — select architectural patterns (workflow, agentic, augmented LLM); choose the simplest tier that meets the needReport a problem with this question

19. 19. An agent loop's cost is dominated by re-sending a long conversation on every turn. An engineer proposes clearing old tool results from the history each turn to cut spend. What should the architect say?

    • A.Clearing old results saves the most when it runs on every single turn, so the schedule should be tightened rather than dropped.
    • B.Clearing old results has no effect until the conversation has already outgrown the window, so the change does nothing now.
    • C.Clearing old results strips the evidence the model relies on, so the loop spends fewer tokens but fails more of its tasks.
    • D.Clearing old results changes the prefix and re-caches the conversation, so it manages the window rather than the bill.✓ Answer

    Context editing removes earlier tool results from the history, which mutates the cached prefix, so the whole conversation must be written to cache again on the next request. Since a cache write costs more than a cache read, a clearing pass can cost more than the tokens it removes. It is a tool for staying inside the context window, not a savings lever; the savings lever for a long loop is a stable cached prefix.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — optimize token usage and cost-performance trade-offs; Anthropic context editing documentation — context editing is a context-window mechanism, distinct from caching economicsReport a problem with this question

20. 20. To cap the spend of a long-running coding agent, an engineer sets a low response token limit. Runs now stop mid-edit and are retried from the beginning. What is the correct configuration?

    • A.Raise the response limit back to a generous backstop and pace the run with a task budget the model itself can see.✓ Answer
    • B.Hold the response limit low and instruct the model in the system prompt to complete each edit within the tokens left.
    • C.Keep the response limit low and lower the reasoning effort, so more of the allowance goes to output than to reasoning.
    • D.Leave the response limit low and add a resume step that continues the run from the point where the last attempt stopped.

    The response token limit is an enforced ceiling the model is never shown, so it is a backstop against runaway output rather than a tuning knob: hitting it is a failed attempt, and the retry costs more than the cap saved. A task budget is the pacing mechanism, because it is advisory and visible to the model, which lets it finish gracefully within the allowance. Prompt instructions cannot give the model information the request does not carry.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — optimize token usage and cost-performance trade-offs; Anthropic task budgets documentation — task budget is advisory and model-visible, distinct from the enforced response ceilingReport a problem with this question

21. 21. A policy assistant begins returning confident answers that quote superseded policy language. The change coincides with a bulk document refresh; the model version, effort setting and latency are all unchanged. Where should the investigation start?

    • A.In the indexing and retrieval step, checking whether the refresh left stale chunks that the retriever still returns as matches.✓ Answer
    • B.In the sampling parameters, checking whether the temperature is high enough for the model to invent authoritative text.
    • C.In the context window, checking whether the refreshed documents are now too large and are being truncated before the call.
    • D.In the model's training data, checking whether the language it memorised predates the version the refreshed corpus holds.

    Confident answers that are wrong in a specific, factual way, with the timing tied to a corpus change and the model configuration unchanged, point at the retrieval layer rather than the model: the assistant is faithfully summarizing whatever chunks it was handed, and those chunks are stale. The first check is whether the refresh reindexed and removed superseded content. Sampling parameters and training data did not change when the behaviour changed.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — diagnose system issues (prompt failure, hallucinations, model mismatch); official sample item on stale retrieval after a document refreshReport a problem with this question

22. 22. A team measured a cheaper model on a random sample of production requests and found it matched the incumbent on 90% of them. They propose switching the whole route. What should the architect require before the cutover?

    • A.Results on the requests the incumbent already answers correctly, since those define the bar the route has to hold.
    • B.Results on a larger random sample, since a few hundred more randomly drawn requests would settle the doubt that remains.
    • C.Results on a sample enriched with the route's hardest requests, since the difficult tail is where retries are paid.✓ Answer
    • D.Results reviewed by the team's own engineers on that same random sample, since review catches what a score misses.

    A random sample is dominated by the median request, but the bill and the failure rate are decided by the hard tail, where a weaker model needs escalations, retries and human handling. Sufficiency must therefore be priced on the hardest tenth of the workload, not on the average case. A larger random draw reproduces the same distribution, human review of easy cases adds little, and testing only what the incumbent already answers correctly hides exactly the failures at issue.

    Source: CCAR-P Exam Guide v1.0, Domain 4 — design evaluation datasets and test frameworks; Anthropic, 'Optimizing for cost and intelligence' — price the hard tail of the workload, not the median caseReport 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. It is delivered via Pearson VUE; Anthropic publishes the current question count, time limit, passing score and fee in the official CCAR-P exam guide. Official certification page →