CCAR-PClaude Certified Architect ProfessionalAnthropicExam UpdateClaude

CCAR-P Exam Changes on December 7, 2026: New Weights, 39 Skills, 70 Questions

Preporato TeamNovember 19, 202611 min readCCAR-P

On December 7, 2026, the Claude Certified Architect - Professional (CCAR-P) exam moves to a new blueprint. The seven domains keep their subjects, but their weights change, Stakeholder Communication and Lifecycle Management halves from 14% to 7%, and the 38 one-line objectives of the current guide are replaced by 39 detailed scenario skills. The exam goes from 63 questions to 70 (60 scored and 10 unscored) in the same 120 minutes. Anthropic published the updated exam guide (version 1.1, effective December 7) on November 17, 2026. This page lists every change and helps you decide which version to take.

The short version

  • Format: 70 questions (60 scored, 10 unscored) in 120 minutes, up from 63. Price ($175) and passing score (720) stay the same.
  • Weights: Integration rises to 21%, Governance to 16%, and Stakeholder Communication falls from 14% to 7%.
  • Skills: 38 objectives become 39 skills, each written as a scenario: evaluate a design, find what fails, pick the control.
  • New topics: deployment surfaces and agent runtimes, Claude Managed Agents, context editing and compaction, LLM-as-judge bias, model version migration, retries and idempotency, prompt injection through MCP and Skills, and Claude Code plugins and permission modes.
  • Before December 7 you sit the current exam. On or after it, prepare from the updated guide.

The format: 70 questions in 120 minutes

CCAR-P before and after December 7, 2026

Until December 6, 2026From December 7, 2026
Exam guideVersion 1.0 (July 2026)Version 1.1 (effective December 7, 2026)
Questions6370: 60 scored, 10 unscored
Time120 minutes120 minutes
Domains77
Item formatMultiple-choice and multiple-response; each item states how many responses to selectScenario-based multiple-choice; each item states how many responses to select
PrerequisitesNoneNone; Claude Certified Architect - Foundations recommended
Passing score720 (scaled 100 to 1,000)720 (scaled 100 to 1,000)
Price$175$175

You get about 1.7 minutes per question instead of 1.9, and CCAR-P stems are long, multi-constraint scenarios, so pacing matters more than before. Ten questions are unscored trial items that look like the rest, so answer every one as if it counts.

The new weights

CCAR-P domain weights, before and after

DomainUntil December 6From December 7Change
Integration19%21%+2
Solution Design and Architecture17%18%+1
Evaluation, Testing, and Optimization16%17%+1
Governance, Safety, and Risk Management14%16%+2
Claude Models, Prompting, and Context Engineering13%13%0
Developer Productivity and Operational Enablement7%8%+1
Stakeholder Communication and Lifecycle Management14%7%-7

The seven freed points go to the technical domains. Stakeholder Communication drops from five objectives to three skills, all about documents an architect produces: an SLA that promises too much, an architecture decision record that hides a trade-off, and a handoff package with a gap. Discovery and expectation management are now tested through those documents, for example by spotting a commitment in discovery notes that the architecture cannot meet.

Every domain's new skills

Solution Design and Architecture (18%)

  • Task Delegation Between Claude and Humans: decide which parts of a business process Claude handles, humans keep, or both share, using reversibility, stakes and accountability.
  • End-to-End Architecture Evaluation: check an architecture against scalability, maintainability and access-control requirements, and find the component that fails first.
  • Architectural Pattern Selection: choose workflow, agentic, augmented-LLM or Skills-based patterns under reuse, governance and risk constraints.
  • Orchestrator and Subagent Decomposition: find where task boundaries, control flow or result synthesis will break against observability and error-recovery requirements.
  • Value-Based Architectural Decisions: pick the option with the best measurable business outcome under a cost or SLA ceiling.
  • Deployment Surface and Agent Runtime Selection: choose between the Anthropic API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry, and between a Messages API loop, a self-hosted Agent SDK or headless Claude Code, and Claude Managed Agents, under residency, compliance and ownership constraints.

Claude Models, Prompting, and Context Engineering (13%)

  • Model Tier Selection: the tier that meets every binding latency, cost and quality constraint, and why each alternative fails.
  • System Prompt Diagnosis and Revision: find the underspecification behind degraded output and constrain it without over-constraining.
  • Prompting Technique Selection: zero-shot, few-shot, prompted reasoning versus native thinking, or structured output, matched to an observed failure.
  • Context Composition and Token Economics: find where tokens can be cut against a budget and call volume without losing what the task needs.
  • Cache Breakpoints and Reusable Packaging: order stable content before dynamic content and package prompts as modules or Skills for reuse.
  • Reasoning Effort and Context Lifecycle: set reasoning effort and pick context editing, the memory tool or compaction for a long-running agent.

Integration (21%)

  • Least-Privilege Tool Configuration: which tool connections to remove, scope down or replace.
  • Authentication and Authorization Posture: credential exposure and privilege escalation risks, and the mitigation for each.
  • Relieving the Binding Performance Constraint: the one change (retrieval depth, model tier, caching, call pattern) that fixes accuracy or latency at least cost to the other.
  • Instrumentation and Monitoring Gaps: the monitoring that would have caught a failure that went undetected.
  • Retrieval Architecture Selection: pre-built RAG, agentic search with tools or full-corpus context, and the contextual-retrieval and fallback settings when you index.
  • Connection Mechanism Selection: MCP, direct API or CLI, or agent-to-agent handoff under latency, security and statefulness constraints.
  • Progressive Discovery vs. Monolithic Context Loading: which context strategy fits the workload's demands and latency tolerance.
  • Production Reliability Controls: rate-limit handling, retry and backoff, idempotency for side-effecting tools, or fallback routing, read from error traces or throughput data.

Evaluation, Testing, and Optimization (17%)

  • Evaluation Coverage Gaps: the missing metric (accuracy, groundedness, latency, cost, safety, security) that would have caught a production failure.
  • Deriving Evaluations from Business Processes: the success criterion and test mix for a model change, prompt change or release gate.
  • LLM-as-Judge Design and Bias: correct position, verbosity, self-preference or judge-drift bias when the judge disagrees with human review.
  • Experimental Design for Configuration Comparison: compare two prompt or model configurations within a budget, including how many trials per case separate the effect from run-to-run variance.
  • Root-Cause Isolation from Logs and Traces: tell prompt failure, retrieval failure, hallucination, tool or subagent error and model mismatch apart.
  • Cost and Latency Optimization: read telemetry, including cache hit rates, and cut spend or latency above a quality floor.
  • Model Version Migration: pin, re-tune prompts, re-baseline evaluations or defer, given a deprecation timeline and regression results.

Governance, Safety, and Risk Management (16%)

  • Guardrail Placement: the control that closes a misuse path without breaking a legitimate use.
  • Prompt Injection Surface and Controls: keep retrieved documents, tool results, MCP server responses and third-party Skills from being treated as instructions.
  • Risk Assessment and Ranking: rank attack surface, failure modes, token and quota exhaustion, and supply-chain risk by exposure.
  • Human-in-the-Loop Checkpoints: where to pause irreversible or high-stakes actions without stalling acceptable-risk paths.
  • Architectural Controls for Compliance Obligations: data residency, retention and zero data retention, PII redaction, audit logging and tenant isolation as evidence for an obligation.
  • Bias Source Identification and Mitigation: tell data bias from prompt-framing and evaluation-design bias, and mitigate at the source.

Stakeholder Communication and Lifecycle Management (7%)

  • Unsupportable Commitments in Requirements and SLAs: the promise the architecture cannot meet, and the clarification or change that fixes it.
  • Architecture Decision Records: revise a draft ADR so its trade-off and residual risk are explicit, without overstating certainty about model behavior.
  • Operational Handoff Completeness: the missing assumption, dependency or runbook element that would block the receiving team.

Developer Productivity and Operational Enablement (8%)

  • Packaging and Distributing Claude Code Workflows: plugins (skills, hooks, subagents, commands, MCP configuration), marketplaces and managed settings, and what to enforce centrally.
  • Agent SDK and Claude Code Configuration: permission modes (default, acceptEdits, plan, auto, dontAsk, bypassPermissions), hooks, subagent definitions, and headless or CI operation.
  • Accountability for AI-Generated Code: the review control that keeps humans owning AI-written code.

What is new for CCAR-P candidates

The current guide names general activities, such as "design multi-agent systems" or "ensure compliance with regulations". The new one names specific products, settings and failure patterns, and almost every skill asks you to evaluate a given design and pick the fix. Topics that are new or now named explicitly:

  • Where Claude runs: Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and Claude Managed Agents as an agent runtime.
  • Long-running agents: reasoning effort, context editing, the memory tool and compaction.
  • Production engineering: rate limits, retries with backoff, idempotency, fallback routing and model version migration.
  • Evaluation rigor: LLM-as-judge biases and how many trials you need to compare two configurations.
  • Security: prompt injection through MCP responses and third-party Skills, quota exhaustion and supply-chain risk.
  • Claude Code at scale: plugins, marketplaces, managed settings and permission modes.

Named regulations (GDPR, HIPAA, FedRAMP) give way to the architectural controls that evidence a compliance obligation, and "ethical AI considerations" becomes finding the source of a measured bias.

Should you test before or after December 7?

  • Nearly ready on the current guide? Book a date before December 7 and sit the 63-question version you studied for.
  • Starting now? Study from the updated guide and book on or after December 7. Put extra time into the newly named topics above, because the current guide does not name them.
  • Strong on stakeholder work? Note that it now carries 7% instead of 14%, and the freed weight moves to Integration, Governance and the other technical domains.

How to prepare for the new version

  1. Read the updated exam guide, including its revised sample questions.
  2. Turn each skill above into a question you can answer: given a design, what fails first and what do you change.
  3. Build or review one production-style Claude system end to end: a deployment surface choice, caching, retries and idempotency, an eval suite with an LLM judge, and a Claude Code setup with hooks and permission modes.
  4. Use the CCAR-P complete guide and the exam domains breakdown, which are being revised to the new guide.
  5. Time yourself with the CCAR-P practice tests. Six follow the current guide for exams until December 6, and 70-question practice tests on the December 7 blueprint are being added next to them; the first went live on November 19, 2026.

Frequently asked questions

Sources:

Free, about 38 minutes

CCAR-P readiness check

Start