Anyone researching the Claude Certified Architect – Foundations certification runs into the same small confusion early: the exam appears under two different codes. Community study repos, forum threads and prep sites — including parts of this one until recently — call it CCA-F. The official Exam Guide calls it CCAR-F.
CCAR-F is the code in the published guide. Version 1.0, effective July 2026, lists it in the exam details table alongside the credential name. It is the string that appears on the official listing and on the booking.
CCA-F is informal shorthand for the same exam. It reads as a natural abbreviation of "Claude Certified Architect – Foundations", it circulated widely in community material, and a great deal of study content still uses it. It is not a different certification, an earlier version, or a lower tier. It is the same sixty-item exam under a name that was never on the paperwork.
There is also a Professional-level Architect exam, CCAR-P, with its own published guide and a separate blueprint. Foundations is the entry point of the two, and the shared "CCAR" stem is the clearest signal that the Architect codes were designed as a family.
Which code to use, and when
The distinction matters in three practical places.
Registering. Use CCAR-F. It is what appears when registering through the Anthropic Partner Academy, with scheduling handled by Pearson VUE. Searching a booking system for CCA-F may return nothing at all.
Searching for study material. Use both, separately. The two codes surface substantially different results, because most existing content was written under one or the other rather than both. A search under a single code returns perhaps half of what exists. Anyone who has downloaded the guide and read the code off it will naturally search CCAR-F — and will miss the large body of community material filed under CCA-F.
Writing or citing. Lead with CCAR-F and mention CCA-F alongside it. Both are true, one is official, and readers arrive with either.
A useful heuristic falls out of this: the code a resource uses is a rough proxy for whether it was written from the guide or from other people’s summaries of it. That is not a reliable rule — plenty of good material predates the v1.0 guide — but it is worth noticing when weighing an unfamiliar source.
What the published guide specifies
Everything below is from Exam Guide v1.0. The guide is the authoritative reference and this is a summary of it, not a substitute — it is worth reading in full before scheduling.
| Credential | Claude Certified Architect – Foundations |
| Exam code | CCAR-F |
| Number of items | 60 |
| Item format | Multiple-choice and multiple-response; each item states how many responses to select |
| Exam structure | 4 scenarios drawn from a bank of 6 |
| Time limit | 120 minutes |
| Passing score | Scaled score of 720 on a range of 100–1,000 |
| Exam fee | $125 USD |
| Validity | 12 months from the date the credential is awarded |
| Delivery | Proctored — online or at a Pearson VUE test centre |
Three of those lines are worth expanding, because each one changes how a candidate should prepare.
Four scenarios appear out of six. All six are published in the guide: a customer support resolution agent, code generation with Claude Code, a multi-agent research system, developer productivity tooling, Claude Code in CI/CD, and structured data extraction. Because only four appear on any given form, core mechanisms recur across them — a candidate meets the same ideas from several angles rather than answering sixty scenario-based items on sixty separate topics. That has a direct consequence for study material: encountering a mechanism more than once is how the exam itself works, not a defect in the practice bank.
Both item formats are used. The guide specifies multiple-choice and multiple-response items, and states that each item tells the candidate how many responses to select. It does not say what share of the exam is multiple-response, so any source quoting a percentage has taken that number from somewhere other than the guide.
The pass mark is not a percentage. It is a scaled score of 720 on a range of 100 to 1,000, which is not the same as answering 72% of items correctly — scaled scoring exists to equate results across exam forms of slightly different difficulty. The $125 registration fee is charged per attempt, which matters more than it first appears.
The five domains and their weights
| Domain | Weight |
|---|---|
| Agentic Architecture & Orchestration | 27% |
| Claude Code Configuration & Workflows | 20% |
| Prompt Engineering & Structured Output | 20% |
| Tool Design & MCP Integration | 18% |
| Context Management & Reliability | 15% |
Because the weights differ by nearly two to one between the largest and smallest, it is worth deciding deliberately how to allocate study time by domain weight rather than working through the blueprint evenly.
Agentic Architecture & Orchestration is the largest single block at more than a quarter of the exam, and it carries the most task statements — seven, covering agentic loops, coordinator and subagent patterns, context passing, enforcement and handoff, hooks, task decomposition, and session management.
The most commonly underestimated is Claude Code Configuration & Workflows. On paper it looks like the domain a working developer absorbs by using the tool. In practice it turns on distinctions nobody stumbles into — when a path-scoped rule beats a directory-level CLAUDE.md, what context: fork actually protects, why plan mode earns its cost on a migration and wastes it on a one-line fix. A fifth of the exam sits there.
The score report gives a pass or fail with a scaled score, plus percent-correct by domain. The domain percentages are informational only; the pass decision rests on the total scaled score alone.
The section most candidates never read
The guide names topics that will not appear on the exam, and several are exactly what an experienced engineer assumes an AI certification must test:
- Fine-tuning or training custom models
- Claude’s internal architecture, training process, or model weights
- Constitutional AI, RLHF, and safety training methodology
- Embedding models and vector database implementation
- Computer use, browser automation, vision and image analysis
- Streaming API implementation and server-sent events
- Rate limiting, quotas, and API pricing calculations
- OAuth, API key rotation, and authentication protocols
- Deploying or hosting MCP servers — infrastructure, networking, containers
- Specific cloud provider configuration
- Token counting and tokenization specifics
- Prompt caching implementation details, beyond knowing it exists
Time spent revising vector stores or working through fine-tuning tutorials buys nothing here. The blueprint tests architectural judgement in production — which mechanism to reach for, and what trade-off comes with it. Not whether a candidate can recall that PostToolUse exists, but whether they know when a hook is the right answer and when a prompt instruction is.
Scheduling, retakes and renewal
Candidates can cancel or reschedule up to twenty-four hours before the appointment; changes inside that window forfeit the fee.
The waiting periods after a failed attempt lengthen each time — fourteen days after the first, thirty after the second, ninety after the third — with a maximum of four attempts per rolling twelve months. The fee applies to each attempt, so a retake costs another $125. That arithmetic is worth doing before booking the first sitting.
The credential lasts twelve months. Renewing on time means reviewing what has changed and completing a free, non-proctored assessment on the Partner Academy, at no fee. A lapsed credential means retaking the full exam at full price. The guide also notes that if exam content changes significantly, holders may be required to retake rather than renew.
Preparing
The guide’s own preparation section is specific and hands-on rather than theoretical: build an agent with the Agent SDK and implement a complete loop with tool calling and session management; configure Claude Code for a real project with a CLAUDE.md hierarchy, path-specific rules and at least one MCP server; design MCP tools with structured error responses; build an extraction pipeline using tool use with JSON schemas; and practise escalation and human-in-the-loop patterns. Section 8 sets out four full exercises with steps.
Practice questions earn their place by surfacing the trade-off reasoning the exam actually tests, which is difficult to rehearse against documentation alone. A free diagnostic gives a rough read on where a candidate stands against the 720 mark, and the practice question bank covers all five domains at their published weights.
This site is an independent practice resource. It is not affiliated with, endorsed by, or operated by Anthropic. To register for the exam, go to the Anthropic Partner Academy and follow the certification links there. Every exam detail above is summarised from the publicly published Exam Guide v1.0 — read the guide itself before scheduling.