Anthropic · CCAR-P

Claude Certified Architect Professional (CCAR-P) Practice Test

Find your weakest domain in 25 questions. Then train on 10 full-length, 120-minute exams of 63 questions, built objective by objective from Anthropic's official exam guide.

Questions: 63
Time: 120 min
Pass mark: 720/ 1,000
Exam fee: $175

Start the free 25-question diagnostic See the 10 full-length exams

By PrepClubs Editorial Team, updated September 26, 2026. Exam facts from Anthropic's official exam guide, v1.0.

Try one now

IntegrationScenario

A team switched its RAG pipeline to a new embedding model. New documents are embedded with the new model into the existing vector index, which still holds millions of chunks embedded with the old model. Queries are embedded with the new model only. Answers drawn from pre-upgrade documents now fail. Retrieval logs show that the correct historical chunks do not appear even in the top 50 candidates. A full backfill that re-embeds the historical archive is already scheduled and will take several weeks. Support needs historical documents to be retrievable starting this week. What should the team do in the meantime?

Similarity between vectors is only meaningful when both vectors come from the same embedding model. New-model query vectors therefore cannot reliably find old-model chunks, which explains why the historical chunks are missing even from the top 50. Keeping the old vectors in their own index lets the team embed each query with both models and search each index with its matching vector. Merging by rank rather than raw score avoids comparing scores across the two spaces. This works right away, while the multi-week backfill that will eventually unify the index is still running.

Why the others are wrong: A reranker helps when relevant chunks are among the retrieved candidates but poorly ordered, yet here the historical chunks never reach the candidate set, so there is nothing for it to reorder. Contextual chunk enrichment improves retrieval when chunks lack surrounding context, but it does not fix the mismatch between query and chunk embedding spaces, and it would require re-embedding anyway. A full re-embed is the correct long-term fix and is already scheduled, but it takes several weeks and so cannot make historical documents retrievable this week.

Source: CCAR-P exam guide, Domain 3

Questions per exam, by official domain weight

Solution Design & Architecture11
Claude Models, Prompting & Context Engineering8
Integration12
Evaluation, Testing & Optimization10
Governance, Safety & Risk Management9
Stakeholder Communication & Lifecycle Management9
Developer Productivity & Operational Enablement4

The exam

CCAR-P at a glance

Exam facts here come from Anthropic's own exam guide, version 1.0, effective July 2026, and its Partner Academy certification FAQ. The exam is for mid- to senior-level architects who take Claude solutions from discovery to production: architecture, integration, evaluation, governance and the stakeholders around them.

Who it is for

Solution architects, AI/ML engineers, technical leads and senior engineers who deliver production LLM systems
Architects who advise clients or internal teams and lead decisions with security, legal and executive stakeholders
People who own a Claude system across its lifecycle, from discovery to monitoring and iteration

Who it is not for

Entry-level developers and casual users of Claude-based apps
Roles limited to isolated tasks such as prompt writing without system design responsibility

Claude Certified Architect Foundations practice tests

Credential: Claude Certified Architect - Professional
Exam code: CCAR-P
Questions: 63 items
Question types: Multiple choice and multiple response. Each multiple-response item tells you how many options to select.
Time limit: 120 minutes, just under 2 minutes per item
Passing score: Scaled 720 on a 100 to 1,000 scale
Delivery: Proctored by Pearson VUE, online or at a test center
Exam fee: $175 USD per attempt. The checkout fee reflects any partner-tier discount.
Validity: 12 months. On-time renewal is a free, non-proctored assessment on the Anthropic Partner Academy.
Score report: Pass or fail, the scaled score, and percent correct by domain
Prerequisites: None required. Anthropic recommends 3+ years in systems architecture or platform engineering and 6+ months of hands-on work with Claude or a comparable LLM system in production.
Eligibility: Currently open only to people at Claude Partner Network organizations. You register with a company email address and must be 18 or older.
Language: English only

The blueprint

Seven domains, turned into the questions you will actually face

Anthropic publishes a weight for each domain. On a 63-item exam that means about 12 questions on Integration and about 4 on Developer Productivity and Operational Enablement. Every PrepClubs exam is built to these counts, give or take one. Below the grid are the 38 objectives the exam items are written against.

63 squares, one per question. Tap a domain to light up its share.

What each domain asks you to do

Domain 1 · about 11 questions

Solution Design & Architecture

Translate business problems into Claude-based AI solutions
Design end-to-end architectures: input, processing, output and feedback loops
Select architectural patterns: workflow, agentic, augmented LLM
Design multi-agent systems and orchestration strategies
Apply decomposition techniques for complex problems
Align solutions to business value: efficiency, transformation, productivity, cost, performance SLAs

Domain 2 · about 8 questions

Claude Models, Prompting & Context Engineering

Select Claude models based on trade-offs
Design system prompts, templates and guardrails
Apply prompting techniques: zero-shot, few-shot, chain-of-thought
Optimize context windows and manage token usage
Implement prompt reuse: caching, modular prompts, Skills

Domain 3 · about 12 questions

Integration

Evaluate tool and agent configuration for capability bloat
Analyze authentication and authorization requirements to find security gaps
Evaluate accuracy-latency trade-offs and justify configuration decisions
Analyze observability challenges and select monitoring strategies at scale
Design a RAG pipeline with appropriate chunking and indexing
Match retrieval strategies to data shape and query pattern
Select the integration mechanism: MCP, API or CLI, agent-to-agent
Evaluate progressive discovery against a monolithic context strategy

Domain 4 · about 10 questions

Evaluation, Testing & Optimization

Define evaluation metrics: accuracy, latency, cost, safety, security
Design evaluation datasets and test frameworks with mixed methods
Run A/B tests and iterative improvements
Diagnose system issues: prompt failure, hallucinations, model mismatch
Optimize token usage, latency and cost-performance trade-offs
Monitor system performance with logging and observability tools

Domain 5 · about 9 questions

Governance, Safety & Risk Management

Implement guardrails and safety controls
Identify risks, limitations and failure modes of LLM systems
Apply human-in-the-loop validation strategies
Ensure compliance with regulations such as GDPR, HIPAA and FedRAMP
Address ethical AI considerations: bias, fairness, transparency

Domain 6 · about 9 questions

Stakeholder Communication & Lifecycle Management

Run structured discovery and requirement gathering
Communicate architectural decisions and trade-offs
Manage stakeholder feedback loops and expectations, including SLAs
Document architectures and provide implementation guidance
Support lifecycle phases: discovery, design, handoff, monitoring, iteration

Domain 7 · about 4 questions

Developer Productivity & Operational Enablement

Configure Claude tools and environments for teams, such as Claude Code
Improve developer workflows with AI-assisted tooling
Support debugging and operational issue resolution

Try it

Three CCAR-P style practice questions

Enterprise scenarios in the shapes the exam uses, including one Select TWO. Answer, then check: each explanation names the Anthropic page or guide objective it relies on.

1 / 3Solution Design & ArchitectureSelect TWO(0 of 2)Multiple response

A legal aid nonprofit is building a business case for a Claude tool that drafts intake summaries. Paralegals review each draft and send the reply. The funder renews the grant only if 90% of applicants receive an initial response within 48 hours. The grant agreement states that an automated acknowledgment does not count as a response: a staff member must review the applicant's situation. Last year the program met the 48-hour window for 72% of applicants. Its cost per applicant served is already below the board's target. Which TWO metrics should the business case lead with? (Select TWO.)

Grant renewal depends on one outcome: 90% of applicants receiving a staff-reviewed response within 48 hours. The program is at 72%, so the business case should lead with that compliance rate. Staff review is mandatory, so paralegal throughput limits compliance. Paralegal review minutes per Claude-drafted summary is the operational metric that shows how the tool closes the gap.

Why the others are wrong: Inquiries resolved without staff involvement do not count as responses under this grant agreement. Deflection rate would lead a case where automated replies satisfied the service requirement. Cost per applicant is already below the board's target, so it is not the gap this case must close. It would lead a case where budget efficiency was the unmet constraint. Satisfaction scores do not measure the funder's renewal condition. They would fit a case where applicant experience or retention was the stated goal.

Source: CCAR-P exam guide, Domain 1

2 / 3Governance, Safety & Risk ManagementSelect TWO(0 of 2)Multiple response

A law firm uses Claude to draft first-pass responses to routine discovery requests. An associate reviews each draft in a review tool that records every edit made and how long each draft stays open. Partners want a monthly early-warning indicator that shows whether associate review is drifting toward rubber-stamping. Firm leadership has set one firm condition: the indicator must be computed only from data the review tool already records, with no changes to the drafts, the workflow, or staffing. Which TWO measurements satisfy this requirement? (Select TWO.)

The deciding detail is that the indicator may use only data the review tool already records, with no changes to drafts, workflow, or staffing. Edit history and open time are both already captured. A falling substantive edit rate is a trend signal of disengagement. Shrinking dwell time per page is a leading indicator of skimming. Neither measure alters the process or adds reviewer effort, so both can feed a monthly early-warning report.

Why the others are wrong: Seeding known errors means altering drafts and running a separate exercise, which the constraint rules out, although it is the strongest way to establish ground-truth detection rates when changing drafts is allowed. Scoring drafts against a labeled set needs data the review tool does not record, and it measures Claude's drafting quality rather than reviewer attentiveness, although it is the right tool for evaluating the drafting step itself. A blind second review needs extra attorney time and a new workflow step, which the constraint forbids, although it is the right tool for confirming whether defects slip through unedited approvals when staffing allows.

Source: CCAR-P exam guide, Domain 4

3 / 3IntegrationScenario

A SaaS company's in-app support assistant sends a large product knowledge base article, followed by the user's short question, on every request. Prompt caching on the article is already enabled, and traces show consistent cache hits and a short time to first token. Nearly all remaining latency and output cost comes from generating answers that average several hundred tokens, but user feedback says only the first paragraph is ever useful. An earlier evaluation found that a smaller model tier answered these questions less accurately, and answer accuracy must not drop. Which change best reduces both total response time and cost?

Generation time and output cost scale with the number of output tokens. The traces show that input processing is already fast and that the excess comes from long answers users do not need. Instructing the model to give a concise one-paragraph answer shortens every response. Lowering max_tokens to match adds a hard ceiling as a backstop. This keeps the same model and the same article, so accuracy is preserved while total time and output cost both fall.

Why the others are wrong: Streaming improves perceived latency but does not shorten total generation or reduce output cost, so it wins only when perceived responsiveness is the sole concern. Caching is already enabled and hitting, so it does nothing about the output-dominated latency here, though it would be the key fix if repeated input processing were the bottleneck. A smaller tier would be faster and cheaper, but the evaluation showed it lowers accuracy, which is disallowed here; it would win where accuracy tolerance allows it.

Source: Prompt caching

Start the free 25-question diagnostic

Six calls a Professional architect is expected to make

Each is grounded in Anthropic's exam guide or docs, and each is a place where the tempting answer treats the symptom.

1
Least privilege means removing capabilities, not watching them.
If a role never needs a tool, take it out of the agent's configuration. Logging and confirmation prompts are compensating controls, not a smaller attack surface.
Source: CCAR-P exam guide, section 8, sample 1
2
Put stable content first if you want caching to pay.
A long system prompt and policy sent on every request should sit before the varying user content so the repeated prefix can be cached, cutting both latency and cost.
Source: Anthropic docs: Prompt caching
3
Confident wrong answers after a data refresh point at retrieval.
When the model and latency have not changed, check the index and the chunks it now returns before touching prompts or models.
Source: CCAR-P exam guide, section 8, sample 3
4
More tools can make an agent worse.
Every extra tool adds selection ambiguity and attack surface. Capability bloat is its own Integration objective on the blueprint.
Source: CCAR-P exam guide, Domain 3
5
An evaluation set is a design artifact, not a sample of logs.
Mix methods: code-graded checks where answers are exact, model grading with a clear rubric where they are not, and human review where stakes are high.
Source: Anthropic docs: Define success criteria and build evaluations
6
An SLA you did not set with the stakeholder is one you will miss.
Latency, accuracy and escalation expectations belong in discovery and in the architecture document, not in a post-launch apology.
Source: CCAR-P exam guide, Domain 6

The course

What you get

full-length exams

63 questions and a 120-minute timer each, the real exam's length and pace.

655

original questions

630 across the paid exams plus the 25-question free diagnostic. No dumps, no recycled items.

domains, blueprint-weighted

Every exam lands within one question of the official weight for each domain.

domain scorecard per exam

Percent correct by domain, the same breakdown Anthropic's score report gives you.

Select TWO items with five options, each stating how many to pick, as the real exam states the count
Enterprise scenarios across financial services, healthcare, retail, government and SaaS
An explanation for every answer that says why each wrong option is wrong
Every explanation ends with the Anthropic docs page or guide objective it rests on
One-time payment, 30-day access, Money-back Guarantee*

Full CCAR-P bank

$69one time, plus tax

10 exams, 630 questions, 30 days of access. No subscription, no auto-renew. Backed by our Money-back Guarantee*.

Start the free 25-question diagnostic Or unlock all 10 exams now

Question quality

Built to fix what practice banks usually get wrong

Before writing a single question we measured the practice banks candidates already use against Anthropic's blueprint and the sample items in its guide. These are the problems we designed out.

Before release, two independent reviewers answered every question blind, without our answer key, and every question we ship matches both. We also hid the questions and showed a reviewer only the answer options: on our bank it picked the right answer about 1 time in 3, close to the 1 in 4 of pure guessing, because every wrong answer is a sound technique that loses only on a detail in the scenario.

Multiple response, rehearsed

Candidates who passed CCAR-P report a large share of multiple-response items, and the best-selling banks we measured had none. About a quarter of our items are Select TWO with five options, scored all or nothing.

Trade-offs, not trivia

Items put you in an enterprise situation with one deciding constraint, such as a compliance boundary, an SLA, a data shape or a stakeholder, and ask for the best call. Every wrong option is a real approach that fails on that constraint.

Weighted to the blueprint

Integration gets about 12 of every 63 questions and Developer Productivity about 4, so your practice time goes where the points are.

No answer-shape tells

Options are written to a similar length and form, and answer positions are balanced across every exam, so the longest or most careful-sounding option is not a shortcut to the key.

All 38 objectives

Every objective in the guide has its own questions, including the ones banks skip: progressive discovery, capability bloat, A/B testing and lifecycle handoff.

Checked against current docs

Every key was checked against Anthropic's docs and the exam guide in September 2026, and a second reviewer answered every question blind before it shipped.

Scoring

How CCAR-P scoring works

The exam is criterion referenced: you are measured against a fixed standard, not against other candidates. Anthropic set the cut score of 720 on a 100 to 1,000 scale through a formal standard-setting study, so a scaled 720 is not the same thing as 72 percent correct, and Anthropic does not publish a raw-score conversion.

Your score report shows pass or fail, the scaled score, and your percent correct in each of the seven domains. Those domain percentages guide your study; the pass decision rests only on the total scaled score. Integration is the largest domain at 19 percent.

Multiple-response items tell you how many options to pick. Anthropic does not say whether they give partial credit, so our practice exams score them all or nothing, the stricter assumption. If you clear our exams comfortably under that rule, partial credit can only help you.

Every exam ends with a domain breakdown

Exam 4 · domain scorecardIllustration

+Solution Design & Architecture82%
+Claude Models, Prompting & Context Engineering75%
-Integration67%
+Evaluation, Testing & Optimization80%
+Governance, Safety & Risk Management78%
+Stakeholder Communication & Lifecycle Management89%
+Developer Productivity & Operational Enablement75%

Weakest by points at stake: Integration. Next: re-read the 14 misses, then exam 5.

Percent correct per domain, like Anthropic's own score report, plus the explanation and docs reference for each miss.

Compare

Which Claude certification is right for you?

Anthropic runs four role-based credentials. Figures are from each exam's official guide and the Partner Academy certification FAQ.

CCAO-F

Associate

You use Claude every day for business work: drafting, research, analysis, Projects and connectors. You do not write code.

CCDV-F

Developer

You build with the Claude API, Claude Code and MCP, and ship apps and agents. You write the code.

CCAR-F

Architect

You design production systems on the Agent SDK, Claude Code, the API and MCP, and own the tradeoffs: orchestration, tools, configuration, structured output and reliability.

CCAR-P

Architect Professional

You lead enterprise Claude deployments end to end: integration, evaluation, governance and compliance, and the conversation with stakeholders. This page prepares you for this exam.

	Associate	Developer	Architect	Architect Professional
Exam code	CCAO-F	CCDV-F	CCAR-F	CCAR-P
Built for	Business professionals using Claude day to day, no coding needed	Engineers shipping apps and agents on the API, Claude Code and MCP	Solution architects designing production Claude systems	Experienced architects leading enterprise deployments
Questions	60	53	60	63
Time	120 min	120 min	120 min	120 min
Pass mark	720 / 1,000	720 / 1,000	720 / 1,000	720 / 1,000
Fee	$99	$125	$125	$175

All four are proctored by Pearson VUE and valid for 12 months. Foundations tests how you build production Claude systems; Professional tests how you run them inside an enterprise.

Plan

How to prepare for the Claude Certified Architect Professional exam

4-week plan

About 5 hours a week, for architects with production LLM experience who are new to parts of the Claude platform.

Week 1
Take the free diagnostic. Read the exam guide and mark each of the 38 objectives you could not defend in a design review today. Refresh the workflow, agent and augmented-LLM patterns and model selection trade-offs.
Week 2
Integration, the biggest domain: tool scoping, auth gaps, RAG chunking and retrieval, MCP versus API or CLI, observability. Sit exams 1 to 3 under the timer and log every miss by domain.
Week 3
Evaluation and governance: metrics, datasets, A/B tests, guardrails, human review, GDPR and HIPAA boundaries. Sit exams 4 to 6.
Week 4
Stakeholder and lifecycle scenarios and Claude Code enablement. Sit exams 7 to 10, re-read the explanation for every miss, and book once two exams in a row clear 80 percent.

2-week sprint

About 10 hours a week, for architects already running Claude in production.

Days 1 to 3
Free diagnostic, then your two weakest domains. If Integration is one of them, start there: it is nearly a fifth of the exam.
Days 4 to 8
One timed exam a day (exams 1 to 5). Review every miss the same evening and open the doc page it points to.
Days 9 to 12
Exams 6 to 9, then a focused pass on Select TWO items, where one wrong pick costs the whole item on our scoring.
Days 13 to 14
Exam 10 as a dress rehearsal at your real exam time of day. Check ID, room and system requirements for online proctoring.

Start the free 25-question diagnostic

Exam day

Pearson VUE exam-day checklist

The rules below are from sections 10 to 13 of Anthropic's exam guide. Read them before you book: a wrong ID or a phone on the desk costs you the $175 fee.

What ID do I need?

A valid, unexpired, government-issued photo ID. The name must match your registration exactly; to correct it, contact [email protected] before you schedule.

What can be on my desk?

Nothing. No notes, books, phones, smart watches, headphones, second monitors or recording devices. If testing online you must stay in view of the webcam for the whole session. Any permitted scratch paper is specified by Pearson VUE.

Can I reschedule?

Yes, up to 48 hours before your appointment, per Anthropic's current certification FAQ (the exam guide still says 24). Changes inside 48 hours forfeit the fee, and a no-show or late arrival past the permitted window also forfeits it.

What if I do not pass?

You can retake after a waiting period: 14 days after the first attempt, 30 after the second, 90 after the third. You can sit the exam up to four times in a rolling 12 months, and each attempt costs the full fee.

What do I agree to before starting?

A confidentiality and non-disclosure agreement: all exam content is Anthropic's confidential property. Disclosing it can void your result, revoke the credential and ban you from future exams.

Can I get accommodations?

Yes, for documented needs. Request them through Pearson VUE and wait for approval before you schedule.

FAQ

Claude Certified Architect Professional FAQs

Can I take the Claude Certified Architect Professional exam as an individual?

Not on your own at the moment. Anthropic's certification FAQ says certification is currently available only to organizations in the Claude Partner Network, and registration needs a partner email address on a recognized company domain. The usual route is through your employer.

What is the Claude Certified Architect Professional certification?

Claude Certified Architect - Professional (exam code CCAR-P, sometimes written CCA-P) is Anthropic's advanced credential for architects who design, build and deliver production-grade AI solutions on Claude: choosing models, architectures and API patterns, integrating Claude into enterprise systems, and building in evaluation, security, compliance and governance.

How many questions are on the CCAR-P exam, and how long is it?

63 questions in 120 minutes, a mix of multiple choice and multiple response. Each multiple-response item states how many options to select.

What is the passing score for CCAR-P?

A scaled score of 720 on a 100 to 1,000 scale. The scale is set by a standard-setting study, so 720 is not the same as 72 percent correct.

How much does the CCAR-P exam cost?

The exam fee is $175 USD per attempt, paid through the Anthropic Partner Academy. The fee shown at checkout reflects any discount your partner tier qualifies for, and each retake costs the full fee again.

Do I need CCAR-F before CCAR-P?

No. The guide lists no mandatory prerequisites, courses or earlier credentials; the recommended experience is guidance only and the credential is awarded on exam performance alone.

How difficult is the Claude Certified Architect Professional exam?

It is aimed at mid- to senior-level practitioners, with 3+ years in systems architecture recommended, and it tests trade-off judgment across seven domains, including governance and stakeholder work that pure engineering prep skips. Multiple-response items raise the bar because both picks must be right.

How long should I study for CCAR-P?

Most architects already running LLM systems in production need 2 to 4 weeks. Take the free diagnostic first: if you score well across domains, the 2-week sprint above is enough; if two or more domains are weak, use the 4-week plan.

What is the difference between CCAR-F and CCAR-P?

Foundations (60 items, $125) tests how you build production Claude systems with the Agent SDK, Claude Code, the API and MCP, framed in 6 published scenarios. Professional (63 items, $175) tests how you run them inside an enterprise: integration, RAG, evaluation, governance and compliance, and stakeholder communication.

How long is the certification valid?

12 months from the date it is awarded. On-time renewal is a free, non-proctored assessment; if it lapses you retake the full exam at the full fee.

Are these real CCAR-P exam questions?

No. Every candidate signs a non-disclosure agreement and real items are Anthropic's confidential property. Our questions are original, written from Anthropic's published exam guide and documentation, and we never use dumps or recalled items.

What do I get if I buy the full bank?

Ten full-length timed exams of 63 questions (630 questions), each weighted to the official blueprint, with an explanation and a docs reference for every answer and a domain scorecard after every exam. It is $69 one time, not a subscription, with 30 days of access and our Money-back Guarantee*.

Is PrepClubs affiliated with Anthropic?

No. PrepClubs is an independent practice-test provider. Claude and Anthropic are trademarks of Anthropic, PBC, used here only to describe the exam.

* See the Money-back Guarantee terms. PrepClubs is not affiliated with or endorsed by Anthropic. Claude and Anthropic are trademarks of Anthropic, PBC.

Find your weakest domain before the exam does

25 questions across all seven domains, timed at the real exam's pace, scored by domain.

Start the free 25-question diagnostic