Skip to content
claude-certified-architect

Overview · CCAR-F

Claude Certified Architect — Foundations

Anthropic's technical certification for solution architects building production applications with Claude (exam code CCAR-F). 60 multiple-choice and multiple-response, scenario-based questions in 120 minutes, proctored and closed-book. Passing score: 720/1000. Candidates answer questions from 4 of 6 randomly selected scenarios. No penalty for guessing.

Part of a four-credential family: Claude Certified Associate — Foundations (CCAO-F), Architect — Foundations (CCAR-F) — this guide, Architect — Professional (CCAR-P), and Developer — Foundations (CCDV-F). Delivered via Pearson VUE (OnVUE online or test center), registered through the Anthropic Partner Academy. Per Exam Guide v1.0 (effective July 2026): $125 per attempt, certification valid 12 months, up to 4 attempts per rolling 12 months (14/30/90-day waits after attempts 1–3). Renewal is free when done on time — review what changed since you certified and pass a non-proctored, open-book assessment (retakes allowed; the new 12 months run from your current expiry); let it lapse and the full exam fee applies again. Access remains gated to the Claude Partner Network (registration needs a partner-domain email); the earlier "$99, first 5,000 partner employees free" beta terms are superseded. Delivery moved to Pearson VUE and digital badging to Credly on 2026-06-30. Partner-tier discounts apply at checkout — 50% for Select, Preferred, and Global Premier partners, 100% for Global Premier partners through 2026-12-31 — and candidates must be at least 18. Cancel or reschedule free of charge up to 48 hours before the appointment (the v1.0 guide's 24 hours is superseded by the live policies page). The FAQ lists the exams and prep content in English and Japanese; Japanese translations of all four v1.0 exam guides were published on 2026-10-08, with unchanged content and weights. Not to be confused with Claude Frontier Academy (announced 2026-10-02): a nominated, in-person program for enterprise engineers that awards Claude Resident Engineer and Claude Frontier Deployed Engineer badges, not a fifth proctored exam.

Target: Solution architect with 6+ months experience with Claude APIs, Agent SDK, Claude Code, and MCP.

Current Model Lineup (October 2026)

The exam is model-agnostic, but scenario questions and code samples assume a current lineup. This guide's samples use claude-opus-5-5, the docs' current default for most workloads; the few samples that rely on forced tool_choice use claude-haiku-5-5, the only current model that still accepts it.

Model ID Context Max output Price (in / out per MTok) Notes
Claude Fable 5.1 claude-fable-5-1 1M 128k $10 / $50 (cache reads $0.25) Most capable widely released model (2026-09-01) — for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short; thinking always on; no forced tool_choice; thinking blocks bound to model + history
Claude Opus 5.5 claude-opus-5-5 1M 128k $4 / $20 (cache reads $0.20) Default for most workloads (2026-09-22): long-running agentic coding and knowledge work; thinking always on and can't be disabled; default effort medium; no forced tool_choice; bound thinking blocks
Claude Sonnet 5.5 claude-sonnet-5-5 1M 128k $2 / $10 (cache reads $0.10) Best speed/intelligence balance (2026-09-28); adaptive thinking on by default — disabled is a 400, send {"type": "between_tools"} instead (at low–high effort); API default effort high but recalibrated, so start agentic coding at medium; no forced tool_choice; bound thinking blocks; gains mid-conversation system messages, per-message effort, and task budgets that Sonnet 5 lacked
Claude Haiku 5.5 claude-haiku-5-5 1M 128k $0.10 / $0.50 up to a 100k-token prompt, $0.50 / $2.50 above (cache reads $0.01 / $0.05) Fastest (2026-10-07) — classification, extraction, routing; adaptive thinking on by default, disabled allowed at high effort or below; default effort medium; still accepts forced tool_choice (the response then has no thinking block); append-only history; no server-side refusal fallback

Claude Mythos 5.1 (claude-mythos-5-1) shares Fable 5.1's specs and pricing and has limited access — organizations verified through one of Anthropic's verification programs, such as the Cyber Verification Program for defensive security work. It is the same underlying model served without the dual-use safety measures Fable 5.1 carries; its safeguards depend on the access program, so still handle refusal there. Legacy but still available: Fable 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 5, Sonnet 4.6, Haiku 4.5. Sonnet 4.5 is deprecated (notice 2026-09-30) and retires on 2026-11-30, with Sonnet 5.5 as the named replacement; Opus 4.1 retired on 2026-08-05. Anthropic publishes retirement floors: Haiku 5.5 not before 2027-10-07, Sonnet 5.5 not before 2027-09-28, Opus 5.5 not before 2027-09-22, Fable 5.1 not before 2027-09-01, Opus 5 not before 2027-07-24, Sonnet 5 not before 2027-06-30. The nearest dates are on legacy models — Haiku 4.5's floor passes on 2026-10-15 and Opus 4.5's on 2026-11-24 — and Anthropic promises at least 60 days' notice before a retirement, so check the deprecations page before committing new work to either.

Model IDs from the 4.6 generation onward are dateless but still pinned snapshots, not evergreen pointers. Query client.models.list() / .retrieve(id) for live max_input_tokens, max_tokens, and capabilities (which now include capabilities.thinking.types and capabilities.server_tools) rather than hardcoding a table like this one.

What the 5.x generation changes for architects (Fable 5.1 introduced three breaking changes on 2026-09-01; Opus 5.5 and Sonnet 5.5 inherited all three, and Haiku 5.5 inherited the third — details in Domain 2 §2.3, Domain 4 §4.7, Domain 5 §5.7):

  1. tool_choice: "any" and {"type": "tool", ...} return 400 on Fable 5.1, Opus 5.5, and Sonnet 5.5 (Opus 5.5 and Sonnet 5.5 apply the same check on count_tokens). Use auto plus an explicit instruction, strict: true, or structured outputs. Haiku 5.5 still accepts forced tool use, but a forced call skips thinking.
  2. Thinking blocks are readable only by some models. Fable 5.1 / Mythos 5.1 read each other's and every earlier model's, and Opus 5.5's on the Claude API; Opus 5.5 reads Opus 5 and earlier Opus / Sonnet / Haiku blocks plus, on the Claude API and Google Cloud, Sonnet 5.5's and Haiku 5.5's; Sonnet 5.5 reads Haiku 5.5's there, but not Opus 5 / 5.5 or any Fable or Mythos block; Haiku 5.5 reads only pre-5 models' (Sonnet 5, Opus 4.8, Haiku 4.5, and earlier). A fallback or router switch to a model that can't read them silently drops them (unbilled) and that model re-plans. Sonnet 5.5 and Haiku 5.5 blocks are also account-bound: replayed through a different, unlinked account they are dropped (organization_binding_mismatch).
  3. Editing, reordering, or removing earlier turns invalidates every later thinking block — on Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 alike (a 400 by default for accounts created on or after 2026-08-31). Harnesses must be append-only: freeze system and tools, move mid-session changes into role: "system" messages (tool definitions included, via inline tools), and trim context server-side (compaction on demand / context editing) rather than on the client.

Opus 5.5 adds three more of its own, none of which break Fable 5.1 code: thinking can't be disabled ({"type": "disabled"} and budget_tokens are a 400 at every effort — lower effort instead); the default effort is medium, one level below Opus 5, so a request that omits effort now runs cheaper and shallower than it did (set it explicitly and re-sweep); and on the Claude API and Google Cloud, computer use only through the computer_toolset_20260801 toolset (computer_20251124 returns 400) — a rule that now covers every 5.5 model. Two behavior changes arrive with no error on Opus 5.5 and Sonnet 5.5: the text Claude writes between tool calls comes back as thinking blocks (empty at the default display), and the refusal categories widen — cyber, bio, frontier_llm, reasoning_extraction, and general_harms on Sonnet 5.5, with Haiku 5.5 running its own classifiers too.

Sonnet 5.5's own break is the thinking switch: disabled returns 400 and {"type": "between_tools"} is the lowest setting — no up-front thinking, only short progress notes between tool calls, accepted at low, medium, or high effort, and it fixes the effort for the rest of the conversation. Haiku 5.5 breaks Haiku 4.5 code in the familiar places: budget_tokens, non-default sampling parameters, and prefill are all 400s, and it uses the newer tokenizer (~30% more tokens for the same text).

Fable 5.1 also requires 30-day data retention (no zero-data-retention orgs unless authorized). Priority Tier is not supported on any current model (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) or on Opus 5 / Sonnet 5 — and new Priority Tier commitments are no longer sold, so plan capacity around rate limits and the Batch tier instead.

Docs (since 2026-07): Anthropic split its documentation. API docs live at platform.claude.com/docs/en/*, Claude Code docs at code.claude.com/docs/en/*; the old docs.anthropic.com URLs still redirect. The SDK was renamed from "Claude Code SDK" to the Claude Agent SDK (pip install claude-agent-sdk, npm install @anthropic-ai/claude-agent-sdk).

Program (Exam Guide v1.0, effective July 2026): four credentials (CCAO-F, CCAR-F, CCAR-P, CCDV-F), proctored via Pearson VUE, registered through the Anthropic Partner Academy. $125 for CCAR-F, 12-month validity — superseding the launch-period "$99 / first 5,000 partner employees free" terms.

Refresh history: CHANGELOG.md.

Exam Domains

Domain Weight Deep Dive
Agentic Architecture & Orchestration 27% domains/d1-agentic-architecture.md
Tool Design & MCP Integration 18% domains/d2-tool-design-mcp.md
Claude Code Configuration & Workflows 20% domains/d3-claude-code-config.md
Prompt Engineering & Structured Output 20% domains/d4-prompt-engineering.md
Context Management & Reliability 15% domains/d5-context-reliability.md

The 6 Exam Scenarios

  1. Customer Support Resolution Agent — Agent SDK, MCP tools, escalation logic
  2. Code Generation with Claude Code — CLAUDE.md, plan mode, slash commands
  3. Multi-Agent Research System — Coordinator-subagent patterns, error handling
  4. Developer Productivity with Claude — Built-in tools, MCP servers, codebase exploration
  5. Claude Code for CI/CD — Structured output, batch API, multi-pass review
  6. Structured Data Extraction — JSON schemas, tool_use, validation-retry

The 10 Critical Anti-Patterns

These show up as wrong answers on the exam:

  1. Parsing natural language for loop termination instead of checking stop_reason
  2. Arbitrary iteration caps as primary stopping mechanism
  3. Prompt-based enforcement for critical business rules (should use hooks)
  4. Self-reported confidence scores for escalation decisions
  5. Sentiment-based escalation (sentiment ≠ complexity)
  6. Generic error messages hiding diagnostic context
  7. Silently suppressing errors (returning empty results as success)
  8. Too many tools per agent (18 vs 4-5 recommended) — see Domain 2 §2.3 for how tool search changes this in the current API
  9. Same-session self-review (retains reasoning context bias)
  10. Aggregate accuracy metrics masking per-document-type failures

Key Decision Frameworks

Decision Option A Option B
Enforcement Programmatic hooks — financial/safety Prompt-based — best-effort
Execution mode Plan mode — multi-file, architectural Direct — single-file, obvious
tool_choice "any" — guarantee tool call Forced — specific tool first. Both return 400 on Fable 5.1 / Mythos 5.1 / Opus 5.5 / Sonnet 5.5 — there, auto + instruction + strict: true, or output_config.format; Haiku 5.5 still accepts both (no thinking on a forced call)
API type Synchronous — blocking workflows Batch — latency-tolerant
Error handling Structured context — category, retry, partial Never generic or silent
Escalation Immediate — explicit customer request Resolve first — within capability
Code review Multi-pass — large PRs Single-pass — small, focused
Context passing Always explicit in subagent prompts Never rely on auto-inheritance
Thinking Adaptive ({"type": "adaptive"}) on every current model budget_tokens is removed — 400 on every current model and on Opus 5 / Sonnet 5 / 4.7 / 4.8; disabled is a 400 on Fable 5.x and Opus 5.5 (lower effort instead) and on Sonnet 5.5 (send between_tools); Haiku 5.5 accepts it at high or below
Effort Start at the model's default and sweep — high on Fable 5.1 / Opus 5, medium on Opus 5.5 and Haiku 5.5 (set it explicitly); Sonnet 5.5 defaults to high but its levels are recalibrated, so start agentic coding at medium; xhigh for the hardest coding/agentic work, and the recommended start on Opus 4.8 / 4.7 low / medium — subagents, mechanical and routine tasks
Tool exposure Load upfront — under ~10 tools Tool search + defer_loading — 10+ tools or multi-server MCP
Context reduction Clear (context editing) — verbose tool results Compact — long conversations that need their narrative; prefer on-demand compaction (you pick the moment, keep recent turns, run it in the background) over threshold compaction
State that must persist Memory tool / files Never rely on the conversation surviving compaction
Agent harness Tool Runner / Agent SDK — you host Managed Agents — Anthropic hosts the loop; the sandbox is Anthropic's cloud or self-hosted on your infrastructure
Conversation history Append-only — freeze system/tools, changes via role: "system" messages (inline tool_addition definitions for tools) Never edit or snip earlier turns on Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5 — it invalidates thinking blocks
Progress UX between tool calls thinking.display: "updates" (beta) or "summarized" — Fable 5.x / Opus 5.5 / Sonnet 5.5 return those notes as thinking blocks (on Sonnet 5.5, between_tools brings them back as visible text) Default "omitted" — the UI goes quiet between tool calls, with no error
Many-agent orchestration (Claude Code, Managed Agents) Subagents / forks — a few delegated tasks per turn Dynamic workflows — a script the runtime executes for dozens to hundreds of agents (Managed Agents: multiagent_20261001, server-run)

4-Week Study Plan

Assumes daily study, 1.5–2 hours per day.

Week Focus Topics Daily Goal
1 Core Architecture Agentic loops, stop_reason, multi-agent orchestration, hooks, session management, tool design, MCP config 1 domain section + practice questions
2 Applied Skills CLAUDE.md hierarchy, plan mode, CI/CD, Batches API, explicit criteria, few-shot, validation-retry, structured output 1 domain section + build a small project using each concept
3 Reliability + Hands-On Context management, escalation, error propagation, provenance, hands-on exercises across all domains Review weak areas + build an end-to-end scenario
4 Exam Prep Practice exams, anti-pattern drills, scenario walkthroughs, timed question sets 1 full practice exam per day + targeted review

Resources

API (platform.claude.com)

Claude Code (code.claude.com)

Background reading