Claude Certified Architect: What the Exam Really Tests

Last updated: November 2026

Claude Certified Architect, Foundations (exam code CCAR-F) is a proctored, 60-question, 120-minute exam on judgment across Claude Code, the Claude Agent SDK, the Claude API and MCP. It costs $125, the pass mark is 720 on a scale of 100 to 1,000, and today only people at Claude Partner Network organizations can sit it. The courses Anthropic lists as preparation are free.

The short version

• It is a judgment exam, not a vocabulary test: every question sits inside a production scenario, and the official sample answers reward the root-cause fix over the clever one.
• Most people searching for it cannot book it yet: registration needs a company email at a partner organization, and Anthropic says customers and non-partners cannot certify at this time.
• The exam guide is the real course: Anthropic calls it the authoritative source for scope, and its hands-on exercises are worth doing whether or not you ever sit the exam.

What is the Claude Certified Architect exam, officially?

It is one of four exams in Anthropic's certification program. Third-party write-ups disagree on details, so everything below comes from official sources, chiefly the certification FAQ on Anthropic Partner Academy, the Architect Foundations exam guide, and Pearson VUE's Anthropic page.

The four exams are Claude Certified Associate at $99, Developer Foundations at $125, Architect Foundations at $125 and Architect Professional at $175. Anthropic's certifications page aims Associate at consultants, sellers and delivery leads, Developer at engineers who build with the API, Claude Code and MCP, and Architect at partners who design Claude solutions end to end. Architect Foundations rose from $99 to $125 for registrations from June 30, 2026. Foundations does not convert into Professional, and Professional has no formal prerequisite. The official code is CCAR-F on both the guide and Pearson's page. If you have seen CCA-F, that is not the code Anthropic uses.

The format, per the guide: 60 items, multiple choice and multiple response, and each item tells you how many answers to pick. The questions hang off four scenarios drawn at random from a published bank of six: a customer support agent, code generation with Claude Code, a multi-agent research system, developer productivity tools, Claude Code in a CI pipeline, and structured data extraction. You get 120 minutes, and the FAQ says to plan for about 135 minutes in the seat once check-in is included. Pearson VUE delivers it online through OnVUE or at a test center, in English only.

Scoring is a scaled 100 to 1,000 with the cut at 720. Your report shows percent correct by domain, but pass or fail is decided on the total. A pass earns a Credly digital badge you can share on LinkedIn. The credential lasts 12 months, and on-time renewal is a free, non-proctored assessment. If you fail, you wait 14 days, then 30, then 90, with a maximum of four attempts in a rolling twelve months and the full fee each time.

The guide describes the intended candidate as "a solution architect who designs and implements production applications with Claude", typically with six months or more of building on the API, the Agent SDK, Claude Code and MCP. That is a practitioner bar, not a beginner one.

What Anthropic has not published: pass rates or how many people hold the credential. Some guides repeat launch-period discounts. The only discounts on the FAQ today are partner-tier ones: 50% off for Select, Preferred and Global Premier partners, and 100% off for Global Premier through December 31, 2026.

On eligibility the FAQ is blunt: "Certification is currently available only to organizations in the Claude Partner Network." You sign in with a company email on a domain in your organization's partner record, you must be 18 or over, and you need government-issued ID. You pay on the Partner Academy's exam page, then book through Pearson VUE. Organizations that deliver Claude for clients can apply through claude.com/partners.

Everyone else can still study. The prep list for this exam is seven free courses: AI Fluency, Building with the Claude API, Claude on Google Cloud, Claude Code in Action, Claude 101, Claude with Amazon Bedrock, and Introduction to Model Context Protocol. All seven also sit in the public Anthropic Academy catalogue. Claude 101 on that list is the everyday-work course. Claude Code 101 is not on the exam list, though Anthropic puts it in the partner Claude Architect learning path, and it is where I would start (how to learn Claude Code has the order). Academy courses end with a quiz and a certificate of completion, which proves you finished a course, not that you passed this exam. There is no official practice exam: the FAQ says it was retired in the move to Pearson VUE. The guide's own preparation advice is hands-on: build an agent with the SDK, configure Claude Code for a real project, design and test MCP tools. If you would rather do that build with someone reviewing it as you go, that is the work in our 1-on-1 Claude Code course.

What do the five domains actually test?

The guide's weights are the approximate share of scored items from each domain:

The out-of-scope list is just as telling. Fine-tuning, embedding models and vector database internals, authentication protocol details, pricing calculations and model benchmark comparisons will not appear. That last one matches what I tell students anyway: benchmarks do not predict your results, your own tests do.

The twelve sample questions in the guide all reward one instinct: find the root cause, then pick the smallest fix that is guaranteed to work. When an agent skips an identity check before a refund 12% of the time, the right answer is a programmatic gate, not firmer wording in the prompt. When it picks the wrong tool, the first fix is a better tool description. Domain 3 overlaps heavily with what I cover in Claude Code best practices that actually matter, especially where instructions should live.

Where does the tool and MCP domain trip people up?

On the difference between MCP and an API, and I have been part of the problem.

A long-standing student asked me on a call this autumn to explain the difference. I answered badly. I said they were basically the same, and described an API as "kind of like tokens". Neither is right, and the mix-up is common enough that I now teach it on purpose.

An API is the interface a service exposes, and your code decides when to call it. A token or key is the credential that lets you in, not the API. An MCP server usually sits in front of one or more APIs and describes its tools to the model. Claude Code's MCP documentation puts it plainly: servers give Claude "access to your tools, databases, and APIs". With MCP, the model reads the tool descriptions and decides what to call. That is why Domain 2 opens where it does: the guide calls tool descriptions "the primary mechanism LLMs use for tool selection."

The same call produced a second mistake. He wanted a bot to sort the email he is copied on all day, and I told him he did not need a bot, he could just connect an MCP. That advice quietly assumed Gmail, and plenty of work mailboxes are not Gmail. For a company Microsoft 365 mailbox, Claude's Microsoft 365 connector needs a Microsoft Entra Global Administrator to grant a one-time consent for the whole tenant, and personal Outlook addresses cannot connect at all. That is not a five-minute job. It is a conversation with IT.

The exam will not ask you about that, because authentication details are out of scope. Real projects stop there anyway. What Domain 2 does test is the judgment next to it: preferring an existing MCP server for a standard integration over building your own, and choosing project scope in .mcp.json for shared team tools against user scope for personal ones. If MCP itself is still fuzzy, start with MCP for beginners.

Why does the reliability domain care whether your system can count?

Because the failure is invisible until someone asks the wrong kind of question.

In October a student and I tested a retrieval system to failure, on a fictional company dataset built for the course. A simple factual lookup passed. A question about something that had not happened yet was correctly refused. Then we asked who had the highest salary, and it answered confidently and wrongly. Nobody noticed at the time it was given.

The cause is structural. Top-k retrieval hands the model a fixed handful of chunks, twenty in this case, and the top earner was not among them. The model ranked what it was given. Counting questions, superlatives and negation ("who has never received an award?") all need the whole dataset, and no prompt wording rescues them. I suggested on the call that more retrieval might fix it. It would not: a bigger k only moves the ceiling.

You will not be asked about chunking, but Domain 5 expects you to know that "aggregate accuracy metrics (e.g., 97% overall) may mask poor performance on specific document types or fields," and Domain 4 expects you to spot when retries are useless because "the required information is simply absent from the source document." Our highest-salary answer was exactly that: the fact was absent from what the model saw. The fixes are the ones the guide rewards. Route aggregation questions to a database query, and tag test questions by type so a passing average cannot hide a failing category. Build your first RAG app walks the pipeline these failures live in.

Frequently Asked Questions

Can anyone become a Claude Certified Architect?

Not at the moment. Only people at Claude Partner Network organizations can register, using a company email on the partner's domain. Anyone can take the free prep courses.

How much does the Claude certification exam cost?

Architect Foundations is $125, up from $99 for registrations from June 30, 2026, and each retake costs the full fee. It is not free unless your company is a Global Premier partner, which gets 100% off through December 31, 2026.

How hard is the Claude Certified Architect exam?

Anthropic publishes no pass rate, so nobody can give you a number. The guide pitches it at someone with six months or more of hands-on building, and says its twelve sample questions "illustrate the format and difficulty level of the exam." Try them cold before you book.

Is there an official Claude certified architect course?

Anthropic offers a prep list of seven free courses and a free Claude Architect learning path on its Partner Academy, but names the exam guide as the authoritative source for what is tested. Read the guide first and use the courses to fill gaps.

Is there an Anthropic Claude Code certification?

Not as a separate exam. Pearson VUE lists four Anthropic exams and none is a Claude Code exam: Claude Code configuration is 20% of Architect Foundations instead. Free courses such as Claude Code 101 end in a certificate of completion, and the Partner Academy has a free Claude Code badge path for consultants, but neither is one of the four certification exams.

Is the Claude certification worth it?

If your company is in the partner network, yes, because the preparation forces you to do real builds. If you are not, the credential is out of reach for now, but the guide and its exercises are free, and I have not seen a clearer map of what production Claude work involves.

Studying for the exam, or building the real thing?

The domains read cleanly on paper and get messy the first time a tool returns an error your agent cannot interpret. If you want help turning the exam guide's exercises into a build that works on your own problem, book a free 15-minute Discovery Call and bring the domain you find hardest.

Written by Michael Murr for AI Tutor Code: private 1-on-1 online tutoring for professionals learning Python, AI tools, Data Science, ML, and LLM engineering. 200+ students taught, 3,000+ hours delivered.