Claude Certification for Developers: CCDV-F vs CCAR-F Path Guide

For a software engineer building with Claude's API today, claude certification for developers means one credential above all others: the Claude Certified Developer, Foundations (CCDV-F), launched 12 March 2026 as part of Anthropic's Claude Partner Network. Which track to choose and how to allocate revision time across eight domains are the two decisions that most determine whether the $125 investment pays off. This guide addresses both directly.

Which Claude certification track should developers choose?

Four tracks are live under the Claude Partner Network, each a proctored Pearson VUE exam. Developers typically compare three:

CCAO-F is scoped to general Claude usage rather than API-level engineering; its blueprint does not test production integration patterns, structured output design, or model-tier selection at technical depth. CCAR-F tests solutions-architect skills: multi-agent orchestration, hub-and-spoke topology, and system-level reliability at scale. Both are coherent credentials for the right candidates, but neither validates what a developer building applications actually does day-to-day.

CCDV-F covers API integration patterns, prompt construction, model selection under constraints, tool wiring, and the security and safety concerns that arise when Claude is embedded in production pipelines. If your work involves shipping features that call the Claude API, CCDV-F is the track that validates those skills.

One structural difference matters for preparation strategy. CCAR-F draws four scenarios at random from a bank of six, meaning which content appears on a given sitting depends partly on chance. CCDV-F has no scenario bank: items are written directly against the domain skill statements. That makes CCDV-F preparation more deterministic. Master all eight domains and you have covered every testable area.

As of 3 June 2026, the Claude Partner Network had more than 40,000 partner applicant firms and over 10,000 certified individuals across all tracks, per Anthropic's public announcements.

What is the CCDV-F exam format?

The CCDV-F consists of 53 items in a 120-minute sitting. Items are multiple-choice and multiple-response; each item states how many answers to select. Delivery is through Pearson VUE, either online-proctored or at a physical test centre.

Scores run on a 100-to-1000 scale, with 720 as the passing threshold. The score report shows pass or fail, the scaled score, and percent-correct by domain. Anthropic does not publish the raw-to-scaled conversion table, so treating any question count as an exact pass mark is unreliable.

The credential is valid for 12 months from the award date.

Which domain dominates the CCDV-F blueprint?

Applications and Integration (Domain 2) carries 33.1% of the exam weight, more than double the second-largest domain. No other area comes close.

Domain 2 covers the practical mechanics of embedding Claude in real systems: synchronous versus streaming API calls, structured output schema design, webhook patterns for asynchronous workflows, conversation state management across turns, and the integration contracts that allow Claude to operate inside larger application architectures. Candidates who have shipped a Claude-powered feature will recognise much of this material, though familiarity with a subset of it does not guarantee coverage of the full domain scope.

The two genuinely low-yield domains are Claude Code (3.1%) and Eval, Testing, and Debugging (2.6%), which together represent just 5.7% of the exam. Spending significant revision time on these at the expense of Domain 2 preparation is the most common strategic error we observe in candidates who narrowly miss the 720 threshold.

What study order produces the highest return on the CCDV-F?

Sequence domains by exam weight, not by numbered order in the blueprint:

1. Applications and Integration (33.1%): Start here. Time invested in Domain 2 produces more scaled-score points than equivalent time anywhere else.
2. Model Selection and Optimisation (16.8%): Cost-latency trade-offs, prompt caching for stable system prompts, batch processing for offline workloads, and model-tier selection for different task profiles.
3. Agents and Workflows (14.7%): Multi-step pipelines, tool-use loops, and agentic orchestration patterns. The Agentic Architecture & Orchestration concept library covers the production patterns that appear in this domain.
4. Prompt and Context Engineering (11.0%): System prompt design, few-shot construction, and context window management. Review the Prompt Engineering & Structured Output material before moving to tool-specific domains; prompt fundamentals compound across the rest of the exam.
5. Tools and MCPs (10.6%): Tool definitions, MCP scoping hierarchy, tool-choice configuration, and error-response design. The Tool Design & MCP Integration library covers the integration patterns the exam rewards.
6. Security and Safety (8.1%): Prompt injection defences, blast-radius limitation, and safe agentic design.
7. Claude Code (3.1%): Revise last.
8. Eval, Testing, and Debugging (2.6%): Scan-level coverage only.

Domains 1 through 5 account for 87.2% of the exam. Allocate at least 80% of your revision time to this range.

How do Tools and MCPs compare with Prompt Engineering as prep priorities?

At 10.6% and 11.0% respectively, Domains 8 and 6 are nearly equal in weight. The question is not which one outranks the other; it is which to study first.

Domain 6 (Prompt and Context Engineering) tests skills that compound across the rest of the exam: choosing instruction formats, constructing few-shot examples that reliably shift model behaviour, managing context window constraints without degrading response quality, and producing structured output that applications can consume programmatically. Gaps in Domain 6 surface as errors in Domain 2 and Domain 1 items as well, because those domains assume you can already write effective prompts.

Domain 8 (Tools and MCPs) tests tool definition design, the MCP scoping hierarchy, the isError flag convention, and how tool-use errors should propagate in multi-agent systems. MCP Server Integration Best Practices covers the production conventions the exam draws on. Finish Domain 6 before moving to Domain 8.

A minimal tool definition the exam might ask you to evaluate:

{
  "name": "get_order_status",
  "description": "Retrieve the current fulfilment status of an order by order ID. Returns status, estimated delivery date, and carrier name. Use for order-tracking queries only; do not call for returns or cancellation requests.",
  "input_schema": {
    "type": "object",
    "properties": {
      "order_id": {
        "type": "string",
        "description": "The unique order identifier (alphanumeric, 8 to 12 characters)."
      }
    },
    "required": ["order_id"]
  }
}

The exam rewards tool descriptions that guide model selection through specificity and constraint, rather than descriptions that merely label what the tool does. Well-scoped descriptions reduce misrouting; that is the reasoning the exam expects candidates to demonstrate.

How does production app design map to the CCDV-F?

Domains 2 and 5 together cover 49.9% of the exam, and both translate directly to production engineering decisions.

Model selection questions in Domain 5 typically present a constraint profile: a latency budget, a cost ceiling, or a task complexity level, then ask which model tier fits. The practical framing: Haiku for high-volume, low-complexity tasks where cost dominates; Sonnet for the mid-tier balance of capability and price; Opus for the highest-complexity reasoning where latency tolerance is high. Prompt caching for stable system prompts and batch processing for offline or non-time-critical workloads are sub-topics in this domain.

Domain 2 integration questions test how Claude is embedded in real systems. Understanding the Messages API Request-Response Cycle is foundational: tracing what happens between an API call and the stop_reason field is something the exam tests both directly and in scenario context. A representative integration pattern:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=1024,
    system="You are a support assistant. Respond as a JSON object with fields: status, message, next_action.",
    messages=[
        {"role": "user", "content": "My order #4821 has not arrived."}
    ]
)

print(response.stop_reason)     # "end_turn" on normal completion
print(response.content[0].text)

Understanding why stop_reason might return "tool_use" instead of "end_turn", and what the application must do before making the next API call, is the class of question Domains 1 and 2 test together. Streaming responses, where the application must consume server-sent events and handle partial content incrementally, are a related sub-topic in Domain 2.

What does Domain 7 (Security and Safety) test on the developer exam?

At 8.1%, Domain 7 is not the highest-weight domain, but it is one where developers who have not read Anthropic's safety documentation often carry unexpected gaps. The domain covers three areas in practice.

Prompt injection defences: identifying injection vectors in user-supplied input, applying input validation and context boundaries, and understanding why user-controlled content must not be trusted to override system-level instructions. The exam rewards defences that are proportionate to the actual threat surface rather than maximally restrictive.

Blast-radius limitation: designing agentic systems with least-privilege tool access, inserting human-approval gates before irreversible actions, and ensuring that a misbehaving sub-agent cannot trigger actions outside its defined scope. The pattern the exam rewards is constraining what each tool can affect, not just what instructions the agent receives.

Safe multi-agent pipeline design: structuring orchestration so that errors in one agent do not cascade silently into downstream agents, and knowing when a system should fail loudly rather than attempt automatic recovery. Developers familiar with Anthropic's guidance on responsible agentic AI will find this domain intuitive; those approaching it fresh should allocate dedicated revision time.

How do you register for the CCDV-F and what does it cost?

The CCDV-F exam costs $125 per attempt. Registration runs through Anthropic's Partner Academy, which routes candidates to Pearson VUE for scheduling. The exam is available online-proctored or at a physical test centre globally. Tiered Claude Partner Network partners receive discounted first attempts per Anthropic's partner programme terms.

For preparation, our platform's CCDV-F adaptive study, practice exams, and Archie tutoring are available today. Practice exams mirror the real format: 53 questions, scored 100 to 1000 with 720 as the passing bar. Our concept library at /concepts currently covers the five CCAR-F Architect domains; a dedicated CCDV-F concept library is not yet live. AI Skill Certs is an independent preparation platform and is not affiliated with or endorsed by Anthropic.