What the CCAR-P exam measures

What you will be able to do

• Name the seven CCAR-P domains and their published weights
• Describe the exam format, length, delivery, fee and validity
• Explain what a scaled score of 720 on a 100-1,000 scale means and what it does not
• Turn domain weights into a defensible allocation of preparation time
• State plainly what this guide is and where the authoritative source lives

The Claude Certified Architect, Professional credential is aimed at the person who has to make a decision and then live with it. Its subject matter is the full arc of a production Claude system: turning a business problem into an architecture, choosing models and prompts, integrating tools and retrieval, evaluating and optimising, governing risk, talking to stakeholders, and enabling the team that maintains the thing afterwards. Most items put you in front of a tradeoff that has no free answer and ask which design you would defend.

That framing should shape how you read the rest of this page. The blueprint is not a syllabus of topics to memorise. It is a list of decisions you are expected to be able to make.

The credential at a glance#

Two rows carry more weight than they look. Sixty-three items in 120 minutes is a shade under two minutes each, and Professional items tend to carry a paragraph of situation before the question, so the real budget is closer to a minute of thinking once reading is paid for. And because every multiple-response item tells you how many responses to select, you are never guessing whether your answer set is complete. You are being asked to rank defensibility, not to collect everything that is merely true.

There is no scenario bank on this exam#

If you have sat or studied for the Foundations exam, unlearn one structural expectation. Foundations builds its paper by drawing four production scenarios from a published bank of six, so a candidate can usefully learn the settings in advance. CCAR-P has no scenario bank. It is organised purely by objective, and items are written directly against the published objective statements.

The practical consequence is that there is no shortcut through familiarity with a fixed cast of systems. Preparation has to be organised the way the blueprint is: domain by domain, objective by objective. It also means the domain mix of your paper is more stable than a Foundations paper's, because it does not depend on which scenarios you happen to draw.

Seven domains, seven weights#

The weights describe how much of the blueprint each domain occupies. If they map evenly onto 63 items, domain 3 is around twelve items and domain 7 around four, but treat that as arithmetic rather than a promise: the blueprint publishes weights, not item counts.

The comparative reading is the useful one. Integration is the largest domain and covers the widest technical surface, from authorisation gaps to chunking strategy to protocol selection, so it is where a shallow pass leaves the most points on the table. Domains 5 and 6 together are 28%, more than integration, and they are the two that engineers most often assume they can improvise. Governance and stakeholder communication have right answers here, not merely tasteful ones. Domain 7 is 7%, roughly four items, and it is the only domain where you can reasonably decide to be competent rather than strong.

What a scaled score of 720 means#

This is the most misread number on the page. 720 out of 1,000 is not 72% correct. The scale is not a percentage wearing different clothes, and the raw proportion of items required is not published, so nobody outside the certification programme can state it. Anyone who quotes you a raw pass percentage is guessing, and the guess cannot be checked, because candidates never see their raw count.

The reason for scaling is equating. Exam forms are assembled from a pool and are not exactly equal in difficulty, so a slightly harder form needs slightly fewer correct answers to represent the same standard. Scaling absorbs that difference and makes 720 mean one thing on every form. The practical consequence is that you cannot finish a practice bank, including this one, and conclude from the percentage alone that you are safely across the line.

What the score report gives you#

Results come back as pass or fail with a scaled score, plus percent-correct broken down by domain. That breakdown is the artefact worth keeping, especially after a fail, because it distinguishes a broad miss from one domain lost badly, and those call for different responses.

Read the per-domain figures against the likely item counts. If domain 7 carries around four scored items, a single wrong answer moves it by roughly twenty-five percentage points, so a low figure there may be one careless item rather than a gap. Domains 1, 3 and 4, with more items behind them, give a steadier signal. Again, that is arithmetic on a published weight rather than a published count, so treat every item estimate in this course as carrying that caveat.

Where 720 came from#

CCAR-P is criterion-referenced, which means every candidate is measured against a fixed standard rather than against the rest of the cohort. No curve, no quota, no benefit or penalty from the company you sat with. The standard came from a standard-setting study, and scaled scoring equates across forms so that 720 carries the same meaning on a harder paper as on an easier one. Taken together, the pass mark describes a level of competence against the published objectives, which is the case for revising against those objectives rather than against a target percentage.

One consequence deserves saying out loud, because the score report actively invites the wrong inference. The per-domain percentages are informational. Pass or fail is decided by the total scaled score alone, so seven figures beside a verdict are not seven gates you had to pass through. Strength across integration and evaluation can carry a weak showing in domain 7, and the reverse holds too. That is not permission to abandon a domain: 14 per cent of the blueprint is a large donation to make when the total is what counts.

Retakes and appeals#

The waiting period lengthens with each failed attempt.

Four attempts per rolling twelve-month period is the ceiling, per exam, and each attempt is charged at the full fee. Because the limits are per exam, a failed CCAR-P leaves you free to register for another exam in the programme in the meantime.

Concerns about the administration or the result go to Pearson VUE support within 14 days of notification, or within 14 days of the exam date where the result itself is the issue. The standard-setting outcome and the content of individual items are not appealable.

Booking and the day itself#

Registration runs through the Anthropic Partner Academy and Pearson VUE, with any partner-tier discount already applied at checkout. Cancel or reschedule at least 24 hours ahead; inside that window, or on a no-show or late arrival, the fee is gone and you book again. Identification is a valid, unexpired, government-issued photo document whose name matches the registration exactly, and accommodations must be requested and approved through Pearson VUE before you schedule. The session is proctored, which online means staying in view of proctor and webcam with a clear workspace and no communication, and which everywhere means no capturing, copying, photographing or reproducing exam content in any form. Misconduct can void a result, revoke a credential and close off future attempts.

Renewal, or a lapse#

Renewing before expiry is a free, non-proctored assessment on the Partner Academy covering what has changed since you certified. Missing that date costs the full exam at the full fee, so the gap between the two paths is about an hour of attention against 175 USD and another 120 minutes under a proctor. Anthropic may also require a full retake instead of the renewal assessment where exam content has shifted significantly, so treat the short path as the normal case rather than a promise.

Twelve months, and why that is deliberate#

The credential is valid for twelve months from award. That is a statement about how fast this material moves rather than about how much of it you retain. Model names, context limits, protocol features and tooling defaults all change on a shorter cycle than most certifications assume, and an architecture that was current when you sat the exam can be a year out of date by the time it expires.

Two habits follow. Learn mechanisms rather than figures, because the mechanism survives the release note. And when a specific number matters to a real design, check the current documentation at docs.claude.com rather than a lesson, this one included.

There are no mandatory prerequisites for this exam. Anyone may book it. That is not the same as saying anyone should, and the next lesson is about the gap between the two: what Professional asks for that Foundations does not, and what that difference means for how you revise.