What is the Claude Certified Architect (CCA)?
On March 12, 2026, Anthropic launched its first official technical credential alongside the Claude Partner Network: Claude Certified Architect – Foundations, exam code CCAR-F (you'll also see it written CCA-F). It's one of four certifications in the program — Associate (CCAO-F, for business users), Developer (CCDV-F, for engineers), Architect – Foundations (CCAR-F, this one), and Architect – Professional (CCAR-P). By June 2026, Anthropic reported more than 10,000 Claude certifications earned across the program.
What makes CCAR-F different from most vendor certs is the format. It's scenario-based, not definitional: instead of asking what the tool_choice parameter does, it drops you into a system — a support agent that's stuck in a loop, a multi-agent pipeline whose subagents are contradicting each other, a Claude Code CI/CD job behaving differently than local runs — and asks what you'd do. It tests architectural judgment under "the system is broken because…" conditions, which is exactly the judgment production work actually demands. That's also what makes it harder to cram for.
Is the Claude Certified Architect worth it?
Honest answer: worth it for a specific person, not for everyone — and the case for it is newer and thinner than for an AWS or Azure cert, because the whole program is months old.
The case for it. It's a first-mover credential issued by the model vendor itself — when hiring managers do start filtering for "can architect Claude-based systems," this is the official signal. The scenario/decision format means passing signals applied judgment rather than flashcard recall, which is rarer among certs. And if you work at a Claude Partner Network firm, certifications count toward the firm's Services Track scoring — some employers will actively want you to hold it.
The case against it. It launched in 2026, so hiring-manager mindshare is still forming — it doesn't yet carry the weight of an AWS Solutions Architect on a resume. The exam is partner-gated and paid. It's vendor-specific in a fast-moving field, with roughly 12-month validity. And there's a documented MCQ-vs-real-skill gap: one candidate publicly reported scoring 1000/1000 on the official practice exam three times, then 590 on the real thing — the practice material simply doesn't match the real scenario format. A multiple-choice pass, even a scenario-based one, still isn't proof you can ship.
Verdict: if you already build on the Claude API or Claude Code professionally and a partner org is in your orbit, the ~$125 and a few weeks of prep are a reasonable bet on an early credential. If you're not in that position, take the free Academy courses (which are most of the educational value) and skip the exam fee for now — the certification will still be there when it carries more market weight. Certifications also move salaries less than shipped systems do; see what actually drives pay in our AI engineer salary breakdown.
Who is the CCA exam for (and who should skip it)?
The target profile is solution architects, AI/ML engineers, and technical consultants who design production systems on the Claude API, Claude Code, the Agent SDK, and MCP. Prep guides suggest roughly 6+ months of hands-on Claude API experience before sitting it, and the pattern in candidate write-ups is consistent: the best predictor of passing is how much you've already built with Claude, not how many hours you studied.
Who should skip or wait:
- • Pure prompt-writers and chat-only Claude users. Roughly 80% of the exam is agent, architecture, and API material — prompt skill alone won't carry you.
- • People brand-new to AI engineering. Build foundations first — start with our AI engineering roadmap — then come back to certification.
- • Anyone not near a partner org. The exam is gated to the Claude Partner Network. Joining is free, but if partner status is irrelevant to your work, the free Academy courses deliver the learning without the fee.
What's on the exam? The 5 domains and their weights
The exam covers five domains. The weights matter for study planning: domains 1 and 2 together are about 47% of the exam — nearly half of your score rides on agent architecture and Claude Code. Questions arrive as 4 extended scenarios drawn from a bank of 6(customer-support agents, code generation with Claude Code, multi-agent research, developer-productivity tools, Claude Code in CI/CD, and structured data extraction), each scenario carrying a cluster of questions.
| Domain | Weight | What it tests |
|---|---|---|
| 1. Agentic Architecture & Orchestration | 27% | Agentic loops driven by stop_reason (continue on tool_use, halt on end_turn), multi-agent coordination and subagent spawning, task decomposition, session state, iteration caps as guardrails |
| 2. Claude Code Configuration & Workflows | 20% | CLAUDE.md hierarchy and precedence (project / user / enterprise), custom skills and slash commands, CI/CD integration |
| 3. Prompt Engineering & Structured Output | 20% | Explicit success criteria, few-shot examples, JSON schema / structured output, validation-retry loops |
| 4. Tool Design & MCP Integration | 18% | Tool descriptions that describe the user-facing function (not the implementation), structured MCP error contracts with recovery hints, the tool_choice parameter |
| 5. Context Management & Reliability | 15% | Rolling-window context management, provenance preservation across long sessions, escalation patterns when an agent can't proceed |
Why each domain matters in practice: agentic orchestration (27%) is where production agents actually fail — loop control via stop_reason rather than trusting the model to say "I'm done," and knowing that subagents don't share context, are the difference between a demo and a system (the same fundamentals as what makes LLMs agentic). Claude Code configuration (20%) tests whether you understand CLAUDE.md precedence and CI/CD behavior — the operational layer most teams learn by getting burned. Prompt engineering and structured output (20%) is about systems that consume model output programmatically, where a malformed JSON response is an outage, not an inconvenience. Tool design and MCP (18%) covers the contract between model and tools — descriptions the model can act on and errors it can recover from. And context management (15%), the smallest domain, is reportedly the one that surprises candidates most — rolling-window management is the topic people say they'd "never heard of" walking in.
Format, cost, and how to register
Caveat before the numbers: this is an independent guide, not affiliated with Anthropic. Some figures below (question count, pricing history, renewal terms) are widely reported by candidates and prep guides rather than published by Anthropic — verify current details on the official Anthropic and Pearson VUE pages before you book.
| Logistics | Detail |
|---|---|
| Exam code | CCAR-F (Claude Certified Architect – Foundations; also written CCA-F) |
| Questions | About 60 multiple-choice, scenario-based* |
| Duration | 120 minutes, single sitting |
| Delivery | Pearson VUE — OnVUE online proctoring or a test center |
| Open book? | No — closed-book, no notes, no docs, no AI tools |
| Scoring | Scaled 100–1000; pass = 720 |
| Price | $125 per attempt (launched at $99; see cost caveats below) |
| Access | Exam gated to Claude Partner Network members (joining is free); prep courses free to anyone |
| Retakes | 14 days after 1st fail, 30 after 2nd, 90 after 3rd; max 4 attempts per rolling 12 months |
| Validity | ~12 months, with free on-time renewal reported |
*On the question count: most sources say about 60 questions; a minority report 77. Anthropic hasn't published an official count, so plan for "about 60" and don't be thrown if your form differs. On price: the current widely-reported price is $125 per attempt — it launched at $99 in March 2026, and while some early articles claimed the exam was "free for the first 5,000 partner employees," that claim is unverified; Anthropic officially states only a discounted first attempt for tiered partners. Treat $125 as current.
How do you register for the CCAR-F exam?
- Join the Claude Partner Network (free) — apply at claude.com/partners.
- Access the Anthropic Partner Academy through your partner account.
- Register and schedule via Pearson VUE — choose OnVUE online proctoring or a physical test center.
- Complete proctoring setup — system check, webcam room scan, government photo ID; closed-book conditions apply for the full sitting.
- Reschedule or cancel up to 24 hours before your appointment if needed.
One recurring point of confusion: Skilljar hosts only the free courses. The proctored exam itself is delivered by Pearson VUE — finishing an Academy course does not certify you.
How do you prepare? What actually worked for candidates
The prep advice below comes from a small sample — early candidate write-ups on Medium plus a handful of prep blogs — so treat it as practitioner signal, not statistics. It's consistent, though, and it contradicts the default "grind flashcards" instinct in one big way.
- Build something real first. The #1 reported predictor of passing. Ship a multi-agent workflow or an MCP server end-to-end — the exam's scenarios are the failure modes you'll have personally hit.
- Drill scenario questions, not definitions. Write ~10 of your own "the system is broken because…" questions per domain and reason through them. That's the exam's actual shape.
- Don't trust the official practice exam. It's ~25 definition-style questions that don't match the real scenario format — the source of the 1000-on-practice, 590-on-real story above. Use it to check terminology, not readiness.
- Weight your study time by domain. Agentic architecture plus Claude Code is ~47% of the exam. Spend your hours there, not evenly.
- Take the free Academy "Agents" and "Tool Use" tracks. Reported as the highest-leverage official material for this exam — and free to anyone, no account gating.
Candidates also converge on five traps the exam repeatedly checks: (1) subagent context isolation — subagents don't share context, so coordination needs explicit handoff; (2) CLAUDE.md precedence across project, user, and enterprise levels; (3) tool-description leakage — descriptions should state the user-facing function, not the implementation; (4) MCP error contracts — structured errors with recovery hints, never raw stack traces; and (5) rolling-window context management, the least-studied domain. If you can reason about all five from experience, you're most of the way there.
Rough timelines from the same write-ups: daily Claude + agent users needed only a few focused sessions; weekly users new to agents needed 2–3 weeks plus one real multi-agent build; chat-only users should plan 5–8 weeks.
Resources for preparing
- • Anthropic Academy — the free official courses; prioritize the Agents and Tool Use tracks
- • Pearson VUE — Anthropic exams — scheduling, OnVUE requirements, test centers
- • Building Effective Agents — Anthropic's canonical agent-architecture essay; domain 1 in prose form
- • Claude Agent SDK docs — the build-something-real starting point
- • Model Context Protocol — for domain 4; build one server end-to-end
- • Community study guide (GitHub) — an organized community-maintained outline
If you're earlier in the journey, our AI engineering roadmap covers the foundations this exam assumes, and what makes LLMs agentic is a gentler on-ramp to the concepts domain 1 tests hardest.
Frequently Asked Questions
Is the Claude Certified Architect exam free? How much does it cost?
The exam is not free — it's widely reported at $125 per attempt (it launched at $99 in March 2026, and Anthropic officially mentions only a discounted first attempt for tiered partners; early 'free for partners' claims are unverified). What IS free: joining the Claude Partner Network, and all of the Anthropic Academy prep courses, which are open to anyone.
How many questions is the CCAR-F exam, and how long is it?
About 60 multiple-choice, scenario-based questions in 120 minutes (a minority of sources report 77 questions — Anthropic hasn't published an official count). You get 4 extended scenarios drawn from a bank of 6, each with a cluster of questions about what you'd do in that system.
What is the passing score for the Claude Certified Architect exam?
720 on a scaled score of 100–1000. Because the score is scaled, that's not a raw 72% — the exact number of questions you need right varies with form difficulty.
Is the Claude Certified Architect certification worth it?
Worth it if you already build on the Claude API or Claude Code professionally, work at (or near) a Claude Partner Network firm, and want a vendor-issued signal of applied agent-architecture judgment. Skip or wait if you're a chat-only Claude user, new to AI engineering, or nowhere near a partner org — the free Academy courses give you most of the learning without the fee.
How long does it take to prepare for the CCA exam?
It depends almost entirely on how much you've built. Daily Claude + agent builders report needing only a few focused review sessions. Weekly users who haven't built multi-agent systems typically need 2–3 weeks plus one real multi-agent build. Chat-only users should budget 5–8 weeks — the exam is about 80% agent/architecture/API material.
Is the Claude Certified Architect exam proctored?
Yes. It's delivered by Pearson VUE, either online via OnVUE (system check, webcam room scan, government photo ID, closed-book conditions) or at a physical test center. Skilljar only hosts the free courses — the exam itself is not on Skilljar.
What is the retake policy if I fail?
Wait 14 days after a first fail, 30 days after a second, and 90 days after a third, with a maximum of 4 attempts in any rolling 12-month period. Each attempt is paid.
Sources
- • Anthropic Academy — official free courses
- • Pearson VUE — Anthropic certification exams
- • Claude Partner Network — official page
- • Anthropic — Building Effective Agents
- • dev.to — The CCA exam: 5 domains, 6 scenarios explainer
- • GitHub — community CCA study guide
- • findskill.ai — how to prepare for the CCA exam (2026)
Independent guide — not affiliated with or endorsed by Anthropic. Exam details change; some figures here are widely reported by candidates and prep guides rather than Anthropic-published. Verify current pricing, format, and policies on the official Anthropic and Pearson VUE pages before booking.