Claude Certified Architect Professional (CCAR-P) Exam Guide (2026)

If you already own the design of a production Claude system, and you have to defend model, integration, and safety choices to the people who fund the work, this is the sitting.

CCAR-P is Anthropic's Architect Professional exam. It is not Architect Foundations. It is not Associate Foundations. It is not Developer Foundations. The official exam guide names the job as designing, building, and delivering production-grade Claude solutions. That includes model and API pattern selection, prompt and context engineering, enterprise integration, evaluation, security, compliance, and governance.

This page covers the current Version 1.0 blueprint, effective July 2026. Weights and task names below are copied from that PDF.

Who this exam is for

Take it if the work already looks like architecture. You translate a business problem into a Claude system. You pick a pattern, a model, and an integration path. You put evaluation and safety in the design before the first demo. You then explain those choices to security, legal, or an executive sponsor.

The official guide names the audience as mid- to senior-level solution architects, AI and ML engineers, technical leads, and senior software engineers. Those people sit between requirements and implementation. They often work in financial services, healthcare, retail, technology, education, or government. The same PDF describes the minimally qualified candidate as someone who has shipped end-to-end AI systems. That work includes RAG, API integration, orchestration, and evaluation. The candidate can talk about cost, latency, performance, safety, and maintainability as tradeoffs, not as slogans.

Anthropic recommends 3 or more years in systems architecture or platform engineering, and 6 or more months with Claude or a comparable LLM system in production. A foundation in modular design, separation of concerns, and scale is also listed. None of that is a gate. The PDF states there are no mandatory prerequisites or required courses. The credential is awarded from exam performance alone.

Skip it if you only chat with Claude or build Projects as a productivity tool. That sitting is Associate Foundations. Skip it if the job is shipping application code against the API with little architecture work. That sitting is Developer Foundations. Skip it if you want the first technical Architect exam, the one built around agent loops, tools, MCP, and Claude Code configuration. That sitting is Architect Foundations.

Skip it if the role is prompt writing with no system design. The official guide excludes entry-level developers, casual users of Claude applications, and purely non-technical roles.

You can sit Professional without holding Foundations. The Partner Academy hub still tells candidates to start at Foundations, then advance to Professional where that path exists. Most people who have not yet designed agents, tools, and Claude Code workflows will get more from Foundations first. People who already run a production Claude program, and who now own RAG, compliance, and stakeholder lifecycle, are in the Professional audience.

How the exam works

The official CCAR-P exam guide, Version 1.0, sets the sitting.

Scoring is criterion-referenced. You pass by meeting a fixed standard, not by beating other candidates. Domain percents on the score report are study feedback. They do not create a second pass bar.

Register and pay on the Anthropic Partner Academy certification page. Checkout shows any partner-tier discount that applies. Pearson then sends the scheduling steps. Pick online proctoring or a test center. The name on the Pearson profile must match the government ID exactly. Ask for a name correction at certifications-support@anthropic.com before you schedule.

You may cancel or reschedule up to 24 hours before the appointment. A change inside 24 hours forfeits the fee. A no-show or a late arrival also forfeits the fee. Failed attempts wait 14 days, then 30, then 90. You may sit the same exam four times in a rolling 12 months. The fee applies each time.

The badge lasts 12 months. On-time renewal is a free, non-proctored Partner Academy assessment. A lapsed credential requires the full exam at the full fee. If the blueprint changes a lot, Anthropic may require the full exam instead of the renewal assessment.

On exam day the workspace stays clear. Notes, phones, secondary monitors, and study materials are out. You accept a confidentiality agreement before the items start. Declining it ends the session with no refund.

Domain weights

The blueprint has seven domains. Weights come from the job task analysis in the official guide.

Integration is the largest slice. If you only polish prompts and skip RAG, auth, and protocol choice, the math is against you. Domain 7 is small. Do not spend a week on Claude Code team setup and starve Domain 3.

Domains 3, 4, and 1 together are about half the scored items. A week that builds one end-to-end system with retrieval, an eval set, and a feedback loop covers those three at once. That is also the prep path the official guide describes.

How it differs from CCAR-F

Architect Foundations and Architect Professional share a role name. They do not share a blueprint.

The Partner Academy hub says to start at Foundations, then advance to Professional where that path exists. The Professional PDF still allows a direct sit. Foundations does not become Professional by renewal.

CCAR-F, from its own Version 1.0 guide, is a 60-item sitting at $125. It uses five domains. A sitting draws 4 scenarios from a published bank of 6. The tasks sit on agentic loops, coordinator and subagent design, hooks, MCP tool interfaces, CLAUDE.md hierarchy, plan mode versus a direct Claude Code run, structured output, and context reliability.

CCAR-P is a 63-item sitting at $175. It uses seven domains. The Professional PDF does not publish a scenario bank. The new weight sits on enterprise integration, evaluation programs, named compliance, stakeholder lifecycle, and business value pillars. Claude Code still appears, but only as team enablement inside a 7 percent domain.

A useful split when you pick a seat:

• If the hard part is whether a hook must block a refund, whether a coordinator should spawn a subagent, or whether plan mode beats a one-function edit, sit Foundations.
• If the hard part is whether the agent has too many tools, whether the RAG index is the first place to look after a document refresh, or whether a control satisfies a named regulation, sit Professional.
• If the hard part is daily Claude use and Projects, sit Associate.

Do not grind a Foundations bank and expect the Professional items to feel the same. The Professional sample items in section 8 already show a different habit. They ask which change removes privilege, which cache layout cuts cost and latency together, and which layer broke after a corpus refresh.

How to study for the live blueprint

The official guide says there is no single required course. Anthropic does not promise that any resource produces a pass. The same section tells you to combine hands-on work with the blueprint, official docs, one end-to-end system, and the sample items.

Week 1. Design and models. For every practice story, write the business problem in one sentence, then name the pattern. Workflow, agentic, or augmented LLM. Say why the other two lose. Decompose the work until each piece has an owner, an input, and an output. Pick a model from a real tradeoff: quality against latency, or capability against cost. Put the stable system prompt and policy first. Enable prompt caching on that prefix. Do not truncate required policy to save tokens. Do not pick the smallest model because it is cheap.

Official docs to keep open this week: the models overview and the prompt engineering guide. Skills belong here too. The blueprint treats Skills as a reuse strategy next to caching and modular prompts.

Week 2. Integration. This is 19 percent. Build one RAG path with chunking and an index you can rebuild. After you refresh documents, query again and see whether answers drift. Practice progressive discovery. Give the model a way to find the next source instead of stuffing the whole corpus into one context. Compare MCP, a direct API or CLI call, and agent-to-agent for the same job. Write down the auth boundary. If a role never issues refunds or deletes accounts, those tools do not belong on the agent.

The official Domain 3 sample is the habit to copy. Support staff only need to read tickets and draft replies. Least privilege removes the refund and delete tools. Logging those tools, or asking for a confirmation, leaves the privilege in place. A larger model does not fix authorization.

Week 3. Evaluation and governance. Define metrics before you change the system. Accuracy, latency, cost, safety, and security are the five the blueprint names. Build a small eval set you can rerun after a prompt or index change. When answers go confidently wrong and the model version did not change, inspect retrieval first. The official Domain 4 sample uses that pattern after a document refresh.

Then place the safety stack. Input checks, output checks, and tool-call authorization. Name the failure mode if a control is missing. If the stem cites GDPR, HIPAA, or FedRAMP, map a named control, an owner, and an evidence artifact. Human review is a design choice for irreversible or high-cost actions, not a sticker you add to every path.

Week 4. Stakeholders and team enablement. Run a short discovery script. Capture the SLA, the reviewer, and the handoff. Draw the architecture so an implementer can build it without a meeting. Domain 6 is 14 percent. Treat documentation and lifecycle as scored work.

Domain 7 is 7 percent. Spend a short block on team Claude Code setup, shared configuration, and how you debug a live miss. Then stop. Extra time here does not recover a weak Integration week.

Work the three official sample items in section 8 for format. They are illustrative. They are not the live bank. After each one, write the constraint in one sentence before you look at the rationale. If you can pick the answer from a product name alone, you are practicing the wrong habit.

The official Partner Academy prep path for this exam walks the same ground: solution design, enterprise integration, the safety stack, evaluations as a gate, stakeholder discovery, and team enablement. Use it as structured reading. Use the PDF as the outline you are scored on.

Where this exam stops

This sitting tests whether you can design and govern a production Claude system. It does not test whether you can chat fluently. It does not test whether you memorized every Claude Code path from Foundations.

Stay on the Professional blueprint.

• Capability bloat is an Integration item. Extra tools that the role never uses are a risk, not a feature.
• A compensating control is not least privilege. Audit logs and confirmation prompts do not remove an unused refund tool.
• A document refresh that produces confident wrong answers points at retrieval and indexing first, not at silent weight changes.
• A static 8,000-token prefix is a caching problem. Truncating required policy, or moving it into few-shot, misses the official rationale.
• Named regulations in the stem want a mapped control. A generic "be careful with data" sentence is not a GDPR, HIPAA, or FedRAMP answer.
• Stakeholder work is scored. A correct architecture that nobody can approve, hand off, or operate still fails the job the PDF describes.
• Domain 7 is real and small. Do not turn the month into a Claude Code workshop.
• Associate Foundations and Developer Foundations are different exams. Their weights do not apply here.
• Academy-style course quizzes are not this Pearson sitting. Study the exam guide PDF.

The official guide also tells you what prep is not. No course, including the Partner Academy path, is required. No resource is guaranteed. If a study note cannot point at a domain and a task in section 6, drop it.

How this maps to CloudFluently

Study with the Claude Certified Architect Professional practice exam sets. Use the Claude Certified Architect Foundations practice exam sets when the gap is agents, tools, MCP, and Claude Code. Use this page for how the Professional sitting is built and how it parts from Foundations. Treat the official PDF as the outline you are scored on.

Official docs for the work Domain 2 and Domain 3 name: models overview, prompt engineering, prompt caching, MCP connector, and Agent Skills.

Official links

Official exam guide: Claude Certified Architect Professional exam guide PDF. Official prep: Architect Professional prep course. Register from Architect Professional certification after you review the guide. The four-exam catalog lives on the Partner certifications hub. Pearson support for this program is on Pearson VUE for Anthropic.