Skip to main content
    Thursday, November 19, 2026

    Claude Certified Architect Prep

    Blueprint-aligned CCAR-F practice — concept cards, coached drills, strict or guided simulation, and a glossary

    Koko teaching Chris, wearing glasses, alongside humanoids and robots. Little Demi teaches a small robot at her two-screen desk.
    KokoAI Academy

    Prepare for the Claude Certified Architect exam

    The Claude Certified Architect – Foundations (CCAR-F) exam tests how you design production AI systems: agentic loops, tool use, prompt engineering, and reliability. Use the Mock exam for coached practice, then take a blueprint-aligned Simulated exam with 60 questions, a 120-minute strict mode, and 4 fictional scenarios drawn from Koko's synthetic pool of 6.

    Independently authored practice content. It does not reproduce official exam questions, imply Anthropic endorsement, or reproduce Anthropic's undisclosed scoring model.

    The five weighted domains
    Exam format
    Length
    60 questions · 120 min

    A timed, proctored sitting. Koko's strict simulation uses the same published count and clock.

    Item format
    Single & multi-response

    The guide states multiple-choice and multiple-response items, with each item stating how many responses to select. Koko forms use 48 single-select and 12 multiple-response (9 select-two, 3 select-three); Anthropic does not publish a fixed live-exam ratio.

    Passing score
    720 of 100–1000

    Criterion-referenced: you are measured against a fixed standard set by subject matter experts, not against other candidates.

    Score report
    Pass/fail + per-domain

    You get a scaled score plus the percentage you answered correctly in each domain. The domain percentages are diagnostic — the pass decision is the total scaled score.

    Delivery
    Pearson VUE

    Registration goes through the Anthropic Partner Academy; the sitting is online-proctored or at a test centre, per program policy.

    Validity
    12 months

    Renew on time with a free, non-proctored assessment on the Partner Academy. Let it lapse and you retake the full exam at full fee.

    Scenarios
    4 of 6

    Koko draws four fictional, synthetic case studies from its six-scenario pool and groups questions under each scenario.

    Unanswered items
    Answer everything

    Koko's reading, not a quoted rule: the current guide states no rule about guessing either way, and under criterion-referenced scaled scoring with no stated penalty a blank cannot score better than an attempt. Confirm against the current exam guide.

    Style
    Closed-book, scenario-based

    Tests documented Claude mechanics and architecture judgment under realistic trade-offs, including configuration and tool-use details.

    Who it's for
    ~6+ months hands-on

    Aimed at people who have shipped production systems with Claude.

    Practice scoring
    100–1000 Koko Practice Index

    720 is Koko's study target on the same scale. It is not an official score prediction or certification outcome.

    Format facts are based on public information and may change — always confirm against Anthropic's official materials before you sit the exam.

    Koko's six fictional, synthetic practice scenarios (a sitting draws 4)
    Northstar Mobile Customer Resolution Agent

    Agentic support, side-effect controls, and reliable recovery

    Vantage SignalGrid: Global Supply-Chain Intelligence Network

    Multi-agent research, evidence reconciliation, and controlled alerts

    ForgeBank Core Platform Modernization

    Claude Code configuration, codebase exploration, and CI review

    OrionWorks Supplier Obligation Intelligence

    Schema-governed extraction, semantic validation, batch processing, and human review

    Vantage Logistics Dispatch Modernization

    Claude Code workflows for a large, safety-sensitive legacy migration

    Forge DevTools MCP Platform

    A secure, efficient developer-productivity integration

    Brought to you by KokoAI Academy — we build production AI systems like the ones this exam is about. This guide is free, and distills the exam's five domains into the design instincts that actually matter in production.

    November platform update

    Reviewed 2026-11-17. These platform changes supplement study guidance. Certification guides and exam requirements remain unchanged.

    November 9: Fable 5.2 selection and cost

    Claude Fable 5.2 (claude-fable-5-2) targets demanding reasoning and long-running agents. Start with the latest Opus model and move up when evaluations justify it. Mythos 5.2 (claude-mythos-5-2) requires access through an Anthropic verification program. Both have a 1M-token context window, 128K maximum output, and always-on adaptive thinking.

    First-party base pricing is $8 input and $40 output per million tokens, down from Fable 5.1's $10/$50. Fable 5.2 fast mode uses speed: fast with no beta header, offers up to 2.5x output throughput, and costs $16/$80. For an illustrative 1M uncached input tokens and 100K output tokens, that is $12 standard or $24 fast, before other charges. Throughput is not an end-to-end latency guarantee. Verify partner-managed platform availability and prices separately.

    Set effort explicitly: low, medium, high, xhigh, or max; the default is high. Asking for no deliberation does not reliably suppress thinking at medium or above. Thinking consumes max_tokens; a small budget can end with stop_reason: max_tokens before visible text. Evaluate effort, output budgets, quality, cost, and latency together. Fable 5.2 has a separate rate-limit pool from Fable 5 and 5.1.

    Official references: model · migration · fast

    Breaking request and replay constraints

    Fable 5.2 and Mythos 5.2 enforce prefix integrity for every account. Editing system instructions, tools, or earlier messages before replaying a later thinking block returns HTTP 400. An identical retry cannot repair it. Preserve the original prefix or remove incompatible thinking when deliberately reconstructing the conversation; never rewrite provider signatures.

    The two 5.2 models can read each other's thinking on supported platforms. Other models, including Fable 5.1 and Opus 5.5, cannot read their blocks; the API drops unreadable blocks when routing. Preserve an authorized original transcript for compatible later turns and evaluate fallback quality with the context actually received. Thinking is also bound to its creating account or a linked account.

    Use computer_toolset_20260801. The older computer_20251124 returns HTTP 400 on these models except on Amazon Bedrock, where the documented legacy path remains supported. A Fable 5.2 or Mythos 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. Forced tool_choice values any and named tool also return HTTP 400.

    Outside CMEK organizations, supported strict tool use or JSON outputs can enforce structure, but auto plus strict does not force invocation or authorize execution. In organizations using customer-managed encryption keys, both strict tool use and JSON outputs are unavailable on these models: omit strict and output_config.format, validate in application code, or select a compatible model.

    Both models require 30-day retention and are unavailable under ZDR unless Anthropic expressly authorizes it. Otherwise first-party requests return HTTP 400. A self-hosted agent service does not override the inference model's retention requirements.

    Official references: model · migration

    November 12: choose the hosting boundary

    Hosted Managed Agents places sandboxes, the service, the agent loop, and state with Anthropic. Self-hosted sandboxes move tool execution to customer infrastructure while Anthropic retains the service, loop, and session history. Full self-hosted Managed Agents moves the API service, loop, PostgreSQL database, workers, sandboxes, and session history to the customer.

    Full self-hosting supports the Claude API, Amazon Bedrock, Google-managed Claude Models, Microsoft Foundry, or a customer model gateway. Existing SDKs and the ant CLI can target the largely shared API. Request access through the Anthropic account team. Existing hosted agents, skills, vault credentials, memory stores, and history do not migrate automatically.

    The customer operates and secures the server, PostgreSQL 16 or later, workers, sandboxes, credentials, upgrades, observability, backups, and trust boundaries. Anthropic cannot remotely access the self-hosted server or its logs. The customer stores transcripts and controls retention and deletion. Model calls still leave the network for the selected inference provider unless the customer gateway changes that route.

    Self-hosted Managed Agents can qualify for ZDR at the service-state layer because transcripts stay with the customer; the inference provider's separate retention terms still apply. Self-hosting does not automatically establish HIPAA eligibility. Document both boundaries and verify contractual eligibility before selecting the architecture.

    Official references: hosting · custody

    November 5: current posture versus audit history

    GET /v1/compliance/organizations/{organization_id}/security_posture reads current security-control results for eligible Claude Enterprise organizations. It reports posture and cannot change settings. The Activity Feed records attestation changes for historical audit evidence and SIEM ingestion; combine it with current results rather than treating either as a substitute for the other.

    Current results require read:compliance_org_data or read:org_audit. With read:compliance_org_data, attester identity additionally requires read:compliance_user_data; read:org_audit can also authorize it. A missing control is not evidence of a pass. An accepted-risk attestation can produce a passed rating even when the underlying control is not implemented; report the attestation and actual control state separately.

    A 404 can mean posture reporting is unavailable or disabled, not that every control passed. Attestation comments are absent from Activity Feed events. Preserve separate approved evidence for risk-acceptance rationale where needed.

    Official references: posture · results

    Professional migration scenario

    A claims agent is moving from Fable 5.1 to 5.2. It edits earlier tools before replay, falls back to Opus 5.5, uses computer_20251124 on the Claude API, and keeps a 5.1 advisor. Its organization requires CMEK and ZDR, and its latency budget assumes a prompt can suppress thinking. What must pass before promotion?

    Decision rationale: Hold promotion until retention is explicitly authorized or a policy-compatible model is selected. Test prefix-preserving replay and account binding; evaluate the fallback without unreadable 5.2 thinking; migrate the computer toolset and loop; select a compatible 5.2 advisor; replace unsupported structured outputs with application validation under CMEK; remove forced selection; sweep explicit effort and token budgets. Rebaseline quality, cost, latency, and the separate capacity pool in a staged canary with rollback criteria. Neither cheaper tokens nor a successful single request proves migration readiness.

    Blueprint-aligned independent practice, written against exam guide v1.0 (July 2026). Koko's scenarios and company facts are fictional and synthetic. This aid does not reproduce official exam questions or Anthropic's undisclosed scoring model, and is not affiliated with or endorsed by Anthropic. Public exam details may change; confirm them against Anthropic's official materials.