LLM Developer Certification: CCDV-F Exam Blueprint 2026

The CCDV-F (Claude Certified Developer, Foundations) is Anthropic's purpose-built llm developer certification for engineers who build production systems on the Claude API. It is not a broad AI fluency credential: it tests whether you can read, reason about, and diagnose real API calls, streaming patterns, tool definitions, and error-handling logic under timed conditions. The distinction matters when choosing where to invest your certification budget.

Four tracks are live in the Claude Partner Network as of 12 March 2026: Claude Certified Associate (CCAO-F, $99), Claude Certified Architect (CCAR-F, $125), Claude Certified Developer (CCDV-F, $125), and Claude Certified Architect Professional (CCAR-P, $175). The CCDV-F is the track engineers reach for first. It targets developers integrating Claude into applications, distinct from the Architect Foundations track (CCAR-F), which focuses on system-level design decisions. This guide covers the exam mechanics, the eight domain weights, the format traps candidates most often underestimate, and the most efficient study sequence.

What are the exact CCDV-F exam mechanics?

53 items, 120-minute time limit, scored on a 100-to-1000 scale with a passing threshold of 720. Those are the fixed numbers, per Anthropic's official CCDV-F exam guide.

Anthropic does not publish the raw-to-scaled conversion formula, so you cannot reliably express 720 as a fixed number of correct answers. Plan for consistent performance across all eight domains rather than targeting a minimum question count. Each attempt costs $125 USD, and the credential is valid for 12 months from the date it is awarded. The exam is delivered online-proctored or at a Pearson VUE test centre.

One structural difference from the Architect track shapes how you prepare: unlike CCAR-F, the CCDV-F has no scenario bank. CCAR-F draws four scenarios at random from a bank of six at each sitting. CCDV-F items are written directly against the skills in each domain, making every item discrete and domain-targeted. Your study approach should reflect that difference: domain-focused practice rather than scenario-style simulation.

Which domains carry the most weight on the CCDV-F?

Applications and Integration (Domain 2) carries 33.1% of the exam, more than double the next-largest domain. It is effectively a standalone exam within the exam, covering the Messages API, streaming responses, vision inputs, file handling, batch processing, error patterns, and rate limits. Candidates who underinvest here rarely pass.

Domain 5 (Model Selection and Optimisation, 16.8%) is the second-largest domain. Expect questions that require matching a use case to the correct Claude model tier based on latency requirements, cost constraints, and capability thresholds. Extended thinking, prompt caching behaviour, and batch-versus-synchronous trade-offs are all in scope.

Domain 1 (Agents and Workflows, 14.7%) pairs naturally with Domain 8 (Tools and MCPs, 10.6%) because tool definitions and MCP configuration are the foundation on which most agent workflows run. Together they account for 25.3% of the exam, making this combined block the third-largest study priority.

The smallest domains are Claude Code (3.1%) and Eval, Testing, and Debugging (2.6%). At 53 total items, even a small domain represents one or two questions. Do not skip them; narrow them to a single focused study pass.

What does a CCDV-F item actually look like?

Every item is scenario-based: it presents a concrete situation and asks for practical judgment rather than recalled definitions. Multiple-choice items have one correct answer. Multiple-response items require selecting two or more, and every item states how many responses to select. The most common format trap is selecting too many answers on a multiple-response item, which voids the question even when an additional selection is defensible.

A Domain 2 item might present a streaming response body and ask you to identify the correct way to handle a partial tool call mid-stream. A Domain 8 item might show a tool definition with an ambiguous description and ask which edit best resolves misrouting. Understanding how tool descriptions function as a selection mechanism is representative of the depth these items probe.

Here is the kind of streaming pattern a Domain 2 item might ask you to reason about:

with client.messages.stream(
    model="claude-opus-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarise this document."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

The exam will not ask you to complete this code. It will ask you to diagnose its behaviour: why a stream terminates before max_tokens, what stop_reason value signals a pending tool call, or when switching to stream=False is the correct production choice. That diagnostic orientation holds across all eight domains.

Domain 1 agent-loop items follow the same pattern. You might be shown a loop that terminates prematurely and asked to identify which stop_reason case the code fails to handle. Agentic loop anti-patterns covers several of the exact failure modes these items surface.

Within Domain 2, the sub-topics that appear most frequently in scenario items are: streaming event types and their ordering, the difference between synchronous and streaming API calls for latency-sensitive use cases, how the Batch API imposes constraints on request size and response retrieval, the four error categories and their correct handling patterns, and how rate limits behave differently across token-level and request-level buckets. Each is independently testable, and Domain 2 items often require reasoning about two or three of them simultaneously.

How does the CCDV-F compare to other AI developer credentials?

The differentiator is depth at the API layer. AWS AI Practitioner, Azure AI Engineer Associate, and similar platform credentials test broad service knowledge across managed offerings. The CCDV-F tests a narrower slice at significantly greater depth: token mechanics, context window behaviour, tool call formatting, and production error patterns at the raw API level.

For teams already shipping on the Claude API, the CCDV-F is the most directly applicable credential available today. For teams evaluating multiple providers or operating at the platform-services layer, broader platform credentials may be the right starting point. A developer who holds the CCDV-F has demonstrated, in a proctored environment, that they can reason correctly about the API mechanics underpinning production Claude systems. The two credential types represent genuinely different kinds of proof and can reasonably be held in sequence.

What is the most efficient study order for CCDV-F?

Work in weight order, with one practical adjustment: treat Domain 1 (Agents and Workflows) and Domain 8 (Tools and MCPs) as a single study block. A candidate who understands how tool calls are formatted, returned, and re-appended within an agent loop will answer Domain 1 questions more accurately than one who treats the two domains in isolation.

Rational study sequence:

1. Domain 2 (33.1%) - Begin with the Messages API request-response cycle, then advance to streaming mechanics, vision and file inputs, batch API constraints, structured error categories, and rate-limit handling. This single domain demands as much study time as Domains 3 through 7 combined.
2. Domain 5 (16.8%) - Model tier selection logic, latency trade-offs, cost optimisation strategies, and when extended thinking improves output quality versus when it increases latency without benefit.
3. Domains 1 and 8 (14.7% + 10.6%) - Agent loop mechanics, tool definition design, tool design and MCP integration patterns including MCP scoping hierarchy and environment variable expansion in config files.
4. Domain 6 (11.0%) - System prompt construction, context management, few-shot example design, prompt engineering and structured output including JSON schema constraints and output format specification.
5. Domain 7 (8.1%) - Prompt injection risks, safety constraints, secure deployment patterns, and input validation at system boundaries.
6. Domains 3 and 4 (3.1% + 2.6%) - Claude Code CLI mechanics including the non-interactive -p flag and CLAUDE.md configuration, followed by eval design and debugging patterns.

For Domain 6, candidates with existing production Claude integrations may need only a targeted review. The techniques map closely to work already completed in shipping applications. Domains 3 and 4 together represent under 6% of the exam; one focused pass through the official documentation is proportionate to their weight.

Does Claude Code appear on the CCDV-F developer exam?

Yes. Domain 3 (Claude Code) carries 3.1% of the exam, representing roughly one to two items. The exam tests developer mechanics that affect integration workflows: how the -p flag enables non-interactive execution, how CLAUDE.md files configure session behaviour, and how custom slash commands extend the tool within team environments.

Domain 8 overlaps at the integration boundary. MCP configuration patterns that Claude Code surfaces in agentic setups fall within Domain 8 scope. Hooks also appear in Domain 1 agentic workflow questions: tool-call interception hooks and PostToolUse hooks for data normalisation are representative of the patterns those items probe.

Do not over-invest in Domain 3. One thorough pass through the Claude Code documentation covering CLI flags, CLAUDE.md configuration, and slash command mechanics is proportionate preparation for its 3.1% exam share.

Is Anthropic Academy content enough to pass the CCDV-F?

Anthropic's free Academy modules provide solid conceptual coverage aligned to the official exam guide and are the correct starting point for any CCDV-F candidate. They are not sufficient on their own for candidates who have not also built directly with the Claude API.

The gap between Academy study and exam readiness is practical pattern recognition. The exam presents plausible wrong answers that reflect real implementation mistakes: choosing streaming when batch processing would reduce cost at scale, misreading a stop_reason field value, or selecting one response too many on a multiple-response item. Repeated exposure to scenario-format questions, scored against the 720 threshold, builds the judgment the exam rewards.

AI Skill Certs' CCDV-F prep is live today: adaptive study sessions, Archie tutoring that guides with graduated hints rather than giving answers directly, and practice exams scored 100 to 1000 across 53 items mirroring the real exam format. The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, concentrating study time on the domains where your knowledge gap is widest rather than material you have already mastered. AI Skill Certs is independent and not affiliated with, endorsed by, or approved by Anthropic.