Updated
Claude Certified Architect Professional: CCAR-P exam guide
CCAR-P is Anthropic's advanced architect exam: 63 questions in 120 minutes, $175 per attempt, a pass at 720 out of 1000. Foundations tests how you build agentic systems. Professional tests how you run them inside an enterprise: integration, evaluation, governance and the conversation with stakeholders. Here is the blueprint domain by domain, three practice questions in its style, and how to prepare.
Claude Certified Architect Professional arrived on 23 July 2026, together with the Associate and Developer exams. It is the only professional-level Claude exam so far. I passed Architect Foundations in July and have not sat Professional. So this post rests on Anthropic's official exam guide for CCAR-P, read against the exam I did take, and I say which of the two I am speaking from.
The exam in numbers
The questions are multiple choice and multiple response, and each one tells you how many answers to select. Pearson VUE proctors the exam online or at a test center. Anthropic's certification FAQ says the exam is available in English and Japanese, and browser translation tools are not allowed. The score report gives your scaled score and the percentage you got right in each domain. Those percentages help you plan a retake. They do not decide the pass, which rests on the total.
The guide recommends three or more years in systems architecture or platform engineering, six months with Claude or a comparable LLM system in production, and experience delivering an end-to-end system from discovery to operation. That is a recommendation. There are no formal prerequisites, Foundations included.
The seven domains
The exam has seven domains. The weights and objectives come from the official CCAR-P exam guide, July 2026 version 1.0, checked on September 29. The names below match the guide so you can compare them directly.
| Domain | Weight | What it covers |
|---|---|---|
| 1 · Solution Design & Architecture | 17% | Turning a business problem into a Claude solution, choosing between a workflow, an agent and an augmented LLM, multi-agent orchestration. |
| 2 · Claude Models, Prompting & Context Engineering | 13% | Model choice on trade-offs, system prompts and guardrails, context windows, prompt caching, modular prompts and Skills. |
| 3 · Integration | 19% | Capability bloat and least privilege, authentication gaps, RAG pipelines and retrieval strategy, MCP versus API or CLI versus agent-to-agent, monitoring at scale. |
| 4 · Evaluation, Testing & Optimization | 16% | Metrics, evaluation datasets, A/B tests, diagnosing hallucinations and model mismatch, the cost and latency trade-off. |
| 5 · Governance, Safety & Risk Management | 14% | Guardrails, failure modes, human-in-the-loop validation, regulations such as GDPR, HIPAA and FedRAMP, bias and transparency. |
| 6 · Stakeholder Communication & Lifecycle Management | 14% | Discovery, explaining trade-offs, expectations and SLAs, documentation, handoff and iteration. |
| 7 · Developer Productivity & Operational Enablement | 7% | Configuring Claude Code and other Claude tools for teams, AI-assisted workflows, operational debugging. |
Integration is the largest domain and includes retrieval design and access control. Together, domains 5 and 6 make up 28% of the exam. Preparing for them means explaining decisions beyond the code.
How Professional differs from Foundations
| Foundations · CCAR-F | Professional · CCAR-P | |
|---|---|---|
| Questions and fee | 60 · $125 | 63 · $175 |
| Structure | 4 scenarios drawn from a bank of 6 | 7 domains, no scenario bank in the guide |
| Center of gravity | Claude Code, Agent SDK, Claude API, MCP | Integration, evaluation, governance, stakeholders |
| Recommended experience | 6+ months building with Claude | 3+ years in architecture, 6+ months with LLMs in production |
| Claude Code configuration | 20% of the exam | part of a 7% domain |
CCAR-P gives more space to integration, evaluation, risk and agreeing requirements with a client. Preparation needs to cover the rationale for a whole solution, followed by how it will be assessed and operated. The Foundations domain list is not enough for that review.
The two overlap on model and prompt choices, tool design and context, and what I learned for Foundations carries over there. Most of that overlap is in the recurring design decisions I collected while preparing for Foundations.
Who should take Professional first
Go straight to Professional if you already lead enterprise deliveries and your week includes security reviews, eval design and steering committees. Start with Foundations if you build agents hands-on but have not yet owned the governance side. If you are unsure which role you are in, four questions that sort the Claude exams by the work you do settle it quickly.
For a partner firm, the tier counts people with a current eligible certificate who also meet the activity requirement. A second certificate for the same person does not increase that count. Choose Professional for the skills a project needs, rather than to count yourself twice toward the partner threshold.
Anthropic admits to its exams only people from Claude Partner Network organizations. Join our collective of AI consultants and sit all four exams through our partner organization.
Three practice questions in the CCAR-P style
Anthropic’s FAQ, checked on September 29, lists no official Professional mock. The old Foundations mock was retired in the move to Pearson. The CCAR-P guide includes three sample questions. I wrote the three exercises below from its scope. Clients, volumes and measured results are hypothetical assumptions, not accounts of my delivery work or live exam questions.
A healthcare client wants an agent that drafts replies to patient messages. Legal requires that no reply reaches a patient before a clinician reviews it, and the client wants to measure how often clinicians change the drafts. Which design meets both requirements?
- A. Instruct the model in the system prompt to mark every draft "for clinician review".
- B. Route every draft to a review queue where a clinician must approve it before sending, and log each edit against the original draft.
- C. Send drafts automatically when the model reports confidence above 0.9, and route the rest to clinicians.
- D. Have clinicians review a random 10% sample of sent replies every week.
Correct: B. The requirement is a hard gate, so it belongs in the workflow, not in the prompt. The queue enforces the review, and logging edits against the draft gives the client its measurement. The confidence threshold lets unreviewed replies through, and a weekly sample checks replies the patient has already read.
Option C is the tempting one, and it fails for a reason worth knowing beyond the exam: a confidence number the model reports about itself is not a calibrated gate.
Three weeks before go-live, the client's head of operations asks for a 99.9% accuracy SLA on an invoice extraction pipeline. Your evaluation set shows 96.4% field-level accuracy, and most errors come from handwritten documents. What should you do first?
- A. Accept the SLA and plan to close the gap with prompt tuning before launch.
- B. Refuse any SLA, because model output cannot be guaranteed.
- C. Present the measured accuracy by document type, propose an SLA per type with human review for handwritten documents, and agree how accuracy will be measured in production.
- D. Switch to the largest available model and commit to 99.9% once it is deployed.
Correct: C. The job here is to align the expectation with the evidence. Splitting the number by document type shows where the risk sits, a review path covers the weak segment, and an agreed measurement method makes the SLA testable. Accepting an unproven number and refusing outright both fail the stakeholder. A larger model still needs to be tested against the agreed accuracy measure.
A bank wants to classify incoming complaints into 12 fixed categories and route each one by a fixed rule per category. Volume is 40,000 complaints a month, and the categories and rules rarely change. Which architecture fits best?
- A. A multi-agent system in which a coordinator hands each complaint to a specialist agent per category.
- B. One model call per complaint with structured output restricted to the 12 categories, followed by deterministic routing in code.
- C. An autonomous agent with tools to read the complaint, look up policies and choose the route on its own.
- D. A separately fine-tuned model for each category.
Correct: B. Here the categories and routing rules are known, so a single call with constrained output offers a simple baseline. The task does not need an agent choosing its next actions freely. Cost, latency and accuracy still need to be measured on representative data; architecture alone does not guarantee better results. A separate model for every category would also need justification and training data.
The Agentic Architect
Practical patterns, case studies, agentic projects taken apart and AI news. Zero spam, once a week.
You'll get one email to confirm. Unsubscribe with one click. Privacy policy
How to prepare
After reading the guide, plan work in every domain. The exercises below suggest evidence you can submit for review. This is my proposed study plan, not an official course syllabus. You can use one small project throughout; it does not require real client data.
| Domain | Exercise goal | What to build and show |
|---|---|---|
| D1 · 17% | Design the solution | Write the client’s requirements and compare a workflow with an agent. Justify your choice using scope, cost, latency and failure handling. |
| D2 · 13% | Choose models and context | Compare two prompt or model variants on the same examples. Keep the results and cases where each variant fails. |
| D3 · 19% | Connect data and systems | Build a small RAG system with access controls. Test whether a user can retrieve a document they are not allowed to read. |
| D4 · 16% | Evaluate quality | Keep tuning data separate from the final test set. Measure quality and latency by document type as well as across the whole set. |
| D5 · 14% | Define operating rules | Write down who approves a high-risk action, what gets logged and how to disable the system. Test the action being blocked. |
| D6 · 14% | Agree on delivery | Prepare a short sponsor memo: pilot results, limits, acceptance criteria, incident ownership and rollback. |
| D7 · 7% | Hand over to the team | Write setup and failure-handling instructions. Give them to another person and check whether they can finish the task without your help. |
PREP PLAN FOR CCAR-P
- 1. Grade yourself against the blueprint. Go through every objective in section 6 of the guide and mark the ones you have never done in a real project. The guide itself suggests this.
- 2. Build one system end to end, with RAG, evaluation and observability. That is the guide's own recommendation, and it covers domains 3 and 4 at once.
- 3. Work through the preparation materials. Open the Professional path in Anthropic Partner Academy and check its current scope. Finish with the guide’s sample questions; course completion is not a CCAR-P exam pass.
- 4. Take a full mock, twice. Do it once in study mode to read every explanation, then once against the 120-minute clock. Go back to the domains where you lost the most.
- 5. Explain trade-offs out loud. Domain 6 rewards the architect who can justify a decision to a sponsor who does not read code.
For that rehearsal, use the free CCAR-P mock from the top of this guide: 63 questions with an explanation for every answer. It follows the guide’s seven domain weights and mixes single-answer and multiple-response questions. Its score uses its own formula, so read it as a map of weak domains, not as a forecast of your result. If you would rather read them first, all 63 CCAR-P questions with answers are on one page, and three are worked through step by step.
Access works the same as for every Claude exam: only through a Claude Partner Network organization. If your employer is not one, I laid out the options for people outside the network and registration requirements.
Members of the Certified AI Consultant Program can take all four Anthropic exams through our partner organization, including CCAR-P. Admission criteria and fees are on the program page.
Frequently asked questions
How many questions are on the Claude Certified Architect Professional exam?
63 multiple-choice and multiple-response questions in 120 minutes. You pass with a scaled score of 720 on a scale from 100 to 1000.
How much does the CCAR-P exam cost?
$175 per attempt, before any partner-tier discount. A retake is charged again, with any applicable partner-tier discount.
Do I need Architect Foundations before Architect Professional?
No. There is no formal prerequisite, and a Foundations certificate does not upgrade into Professional. They are separate exams with separate fees.
Is there an official CCAR-P practice exam?
No. The only official practice exam, for Architect Foundations, was retired in the move to Pearson on 30 June 2026, before Professional launched. The official exam guide includes three sample questions with answers and explanations. An unofficial 63-question mock written from that guide is free on this site.
What are the CCAR-P exam domains?
Solution Design and Architecture 17%, Claude Models, Prompting and Context Engineering 13%, Integration 19%, Evaluation, Testing and Optimization 16%, Governance, Safety and Risk Management 14%, Stakeholder Communication and Lifecycle Management 14%, and Developer Productivity and Operational Enablement 7%.
Disclaimer. This post is independent and unofficial. It is not affiliated with, authorized or endorsed by Anthropic. Exam facts come from the Claude Certified Architect Professional exam guide (version 1.0, effective July 2026) and Anthropic's certification FAQ, checked on 29 September 2026. The three practice questions are original and are not drawn from the exam. Claude and Claude Certified Architect are names owned by Anthropic.
