GenAI Certification for Developers: CCDV-F Exam Format Decoded

Choosing the right genai certification for developers in 2026 means understanding not just which credential to pursue, but what the exam actually tests and how hands-on preparation translates to a passing score. The Claude Certified Developer, Foundations exam (CCDV-F) is Anthropic's proctored qualification for developers building production applications with Claude. It is scenario-based throughout, rewards engineering judgment over memorisation, and maps closely to the decisions developers make when shipping real Claude-powered systems.

What is the CCDV-F and who should take it?

CCDV-F is the developer track of Anthropic's Foundations certification programme, delivered via Pearson VUE as an online-proctored or test-centre exam. It costs $125 per attempt, runs for 120 minutes, and covers 53 items across 8 weighted domains. The passing score is 720 on a 100 to 1000 scale.

The exam sits within the Claude Partner Network, Anthropic's $100 million programme. As of 3 June 2026, more than 40,000 partner applicant firms had applied and more than 10,000 individuals had earned a Claude certification. CCDV-F is designed for developers who integrate the Claude API, design tools and agents, and evaluate model behaviour in production. Every item is grounded in a concrete engineering scenario; nothing tests pure recall of documentation.

One key structural difference from the architect track is that CCAR-F draws 4 scenarios at random from a bank of 6 at each sitting, so each candidate sees a different sub-set of content. CCDV-F has no scenario bank. Items are written directly against the skills in each domain, which means every domain appears at every sitting.

How is the CCDV-F scored and what does 720 mean in practice?

CCDV-F is scored on a 100 to 1000 scale with 720 as the passing threshold. Anthropic does not publish the raw-to-scaled conversion, so no authoritative figure exists for the number of correctly answered items that produces a 720.

The score report you receive shows pass or fail, the scaled score, and percent-correct by domain. That domain breakdown is the most useful diagnostic data available, particularly for planning a re-sit. Because the scaled score aggregates across all 8 domains, a developer who excels in Applications and Integration but struggles with Model Selection may still fall short of 720 if the deficits are significant enough elsewhere. The credential is valid for 12 months from the award date.

What does the 8-domain structure look like and which areas carry the most weight?

The eight domains and their precise exam weights, sourced from the official CCDV-F exam guide, are:

Domain 2, Applications and Integration, accounts for roughly one-third of the exam at 33.1%. Performance in this domain effectively sets a floor for your scaled score. Items cover API integration patterns, streaming, error handling, and the mechanics of building reliable interactions with the Messages API.

Domain 5, Model Selection and Optimisation (16.8%), tests whether developers can choose the right model for a given cost and latency constraint and avoid over-engineering solutions with large models when smaller ones suffice. Together, Domains 2 and 5 represent 49.9% of the exam.

Domain 1, Agents and Workflows (14.7%), is where agentic architecture concepts appear: structuring multi-step pipelines, managing state between turns, and avoiding loop anti-patterns that cause runaway or stalled agents.

Domains 3 and 4, Claude Code and Eval/Testing, carry a combined weight of 5.7%. They still appear in every sitting and a weak domain-level result affects the score report, but they should not absorb a disproportionate share of preparation time.

What does "multiple-response" mean on the CCDV-F?

CCDV-F items are either multiple-choice (select one correct answer) or multiple-response (select more than one). Per the official exam guide:

You are never left to guess how many answers a multiple-response item requires. The failure mode we see most often is developers selecting a partially correct answer and stopping, treating a multiple-response item as if it were multiple-choice. On a three-answer item, selecting only two is a guaranteed partial miss regardless of which two answers you chose.

The practical implication: before answering, note the number of required responses and commit to finding that many. If you can only identify two of three required correct options, eliminate the obviously wrong choices and make a considered selection rather than leaving a blank.

What hands-on experience actually prepares you for CCDV-F?

The exam does not ask you to recall parameter signatures. It asks you to diagnose what is wrong in a described engineering situation and select the correct response. Preparation is therefore most effective when built around building and debugging real systems rather than passive reading.

Four areas translate directly to exam readiness:

Streaming API integration. Domain 2 presents integration scenarios involving how to handle partial delta events, reconnection logic, and the semantics of stop_reason. A developer who has implemented streaming in production will recognise the failure modes immediately. The typical streaming pattern looks like this:

import anthropic

client = anthropic.Anthropic()

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

    message = stream.get_final_message()
    # stop_reason: "end_turn" | "max_tokens" | "tool_use" | "stop_sequence"

Understanding when each stop_reason value appears and what the correct application response is in each case is a live exam topic.

Tool use and MCP server design. Tool Design and MCP Integration underpins Domain 8 and feeds into Domain 2. The exam tests whether developers understand when to split a tool versus combine it, how tool descriptions act as a routing mechanism, and how the isError flag differs from a valid empty result.

Prompt and context engineering. Domain 6, Prompt and Context Engineering (11.0%), rewards developers who have iterated on prompts in production. Knowing when few-shot examples outperform chain-of-thought, when structured output schemas are necessary, and how context window usage affects reliability are all live topics. Our prompt engineering concept library maps the atomic skills in this domain systematically.

Evaluation pipelines. Domain 4 carries 2.6% of the exam but its questions are specific. A developer who has written even a basic LLM-as-judge evaluation or a regression suite around prompt changes will recognise the terminology and trade-offs without needing to study theory from scratch.

How does Claude Code fit into CCDV-F preparation?

Domain 3, Claude Code, accounts for 3.1% of the exam. The questions are precise: they test configuration hierarchy, how hooks interact with tool calls, and when Claude Code is the appropriate tool versus when a direct API call is preferable.

The distinction matters. Claude Code is an agentic coding environment designed for IDE-native, codebase-aware tasks where the developer works interactively alongside the model. The Messages API is appropriate for programmatic, non-interactive pipelines where the application controls all inputs and outputs. Exam items will present a scenario and ask which approach fits; the answer hinges on whether the task is interactive and codebase-aware or programmatic and latency-sensitive.

MCP spans Domains 2 and 8. A developer who has built and deployed an MCP server, observed how tool descriptions influence routing decisions, and debugged a misrouted tool call has the tacit knowledge these items require. The full concept library at /concepts covers MCP scoping and integration patterns in detail.

How does Domain 7, Security and Safety, appear in practice?

Domain 7 carries 8.1% of the exam. Questions are not abstract policy theory; they are scenario-based decisions about input validation, safe failure modes, and handling model outputs that may be harmful or unexpected.

The exam consistently rewards proportionate responses. A question about handling a potentially harmful tool invocation will typically offer one answer that overreacts (blocking all tool use) and one that underreacts (proceeding without validation). The correct answer applies a targeted check at the right system boundary. Developers who think about validation at entry points rather than attempting to filter inside the model will find this pattern familiar.

Prompt injection is a tested concept. The exam distinguishes between injection arriving through user input and injection arriving through tool results, and asks what mitigation is appropriate in each case. Both attack surfaces are treated as live engineering concerns, not edge cases.

How should you allocate study time across the 8 domains?

A weight-proportional baseline, adjusted for your measured skill gaps, is the most defensible approach. The following allocation assumes a 40-hour preparation block:

If you already have production experience with streaming and API error handling, redistribute Domain 2 hours toward areas where your practice exam score report shows the most gap. The platform's adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, surfacing genuine knowledge gaps rather than topics that feel unfamiliar but are already well understood.

CCDV-F adaptive study, Archie tutoring, and practice exams scored on the same 100 to 1000 scale as the real exam are available on the platform today. The CCDV-F concept library is not yet live; the CCAR-F concept library at /concepts covers 174 atomic concepts mapped to the architect track domains.

What format should you expect on exam day?

CCDV-F is delivered online-proctored via Pearson VUE or at a test centre. The 120-minute window for 53 items gives you approximately 2 minutes and 15 seconds per item. That is workable on straightforward single-answer items but tight on multi-response questions that require evaluating several option combinations carefully.

There is no scenario bank, which means every domain appears in every sitting. A candidate cannot benefit from a draw that happens to favour their strongest domain. Consistent coverage of all 8 domains is the only reliable strategy.

After the exam, the score report shows pass or fail, the scaled score, and percent-correct by domain. Combined with practice exam feedback taken under timed conditions, that breakdown gives a clear picture of where additional preparation is needed before a second attempt.