On December 7, 2026, the Claude Certified Developer - Foundations (CCDV-F) exam changes in three ways. It grows from 53 to 70 questions (60 scored and 10 unscored) in the same 120 minutes, its domain weights even out so that Applications and Integration drops from 33.1% to 20%, and every domain's skill list is rewritten around scenarios you analyze and fix. Anthropic published the updated exam guide (version 1.1, effective December 7) on November 17, 2026. This page lists every change, from the format to each new skill, and helps you decide which version to sit.
The short version
- Format: 70 questions (60 scored, 10 unscored) in 120 minutes, up from 53. Price ($125) and passing score (720) stay the same.
- Weights: Applications and Integration 33.1% to 20%; Eval, Testing, and Debugging 2.6% to 10%; Claude Code 3.1% to 8%; Prompt and Context Engineering 11.0% to 15%.
- Skills: all 25 old skills are replaced by 24 new ones, written as tasks such as diagnosing a misbehaving agent or fixing failing integration code.
- Before December 7 you sit the current exam. On or after it, prepare from the updated guide.
The format: 70 questions in 120 minutes
CCDV-F before and after December 7, 2026
| Until December 6, 2026 | From December 7, 2026 | |
|---|---|---|
| Exam guide | Version 1.0 (July 2026) | Version 1.1 (effective December 7, 2026) |
| Questions | 53, all scored | 70: 60 scored, 10 unscored |
| Time | 120 minutes | 120 minutes |
| Item format | Multiple-choice and multiple-response; each item states how many responses to select | Scenario-based multiple-choice; each item states how many responses to select |
| Passing score | 720 (scaled 100 to 1,000) | 720 (scaled 100 to 1,000) |
| Price | $125 | $125 |
| Languages | English and Japanese | English and Japanese |
Two practical effects follow. First, the pace tightens: 53 questions in 120 minutes left about 2.3 minutes each, and 70 leaves about 1.7. Second, ten questions are unscored trial items that you cannot identify, so every question still deserves a full answer. The guide also describes every item as scenario-based, and it still says each item states how many responses to select, so expect questions that ask for more than one answer.
Try 15 CCDV-F practice questions free
They come from the CCDV-F practice tests, and each option shows its explanation once you answer.
- 6 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
The new domain weights
CCDV-F domain weights, before and after
| Domain | Until December 6 | From December 7 | Change |
|---|---|---|---|
| Applications and Integration | 33.1% | 20% | Down 13.1 points |
| Agents and Workflows | 14.7% | 17% | Up 2.3 |
| Prompt and Context Engineering | 11.0% | 15% | Up 4.0 |
| Model Selection and Optimization | 16.8% | 12% | Down 4.8 |
| Eval, Testing, and Debugging | 2.6% | 10% | Up 7.4 |
| Tools and MCPs / Tools and MCP | 10.6% | 10% | About the same |
| Claude Code | 3.1% | 8% | Up 4.9 |
| Security and Safety | 8.1% | 8% | About the same |
The domain names stay the same, apart from "Tools and MCPs" becoming "Tools and MCP". What moves is the balance. The current exam is a third API integration, and evaluation plus Claude Code together are under 6%. The December 7 exam spreads the weight: no domain is above 20%, and evaluation, Claude Code and prompt and context engineering together grow from 16.7% to 33%. On 60 scored questions, Eval, Testing, and Debugging goes from one or two questions to about six.
Every domain's new skills
The guide lists the skills each domain's items are written against. Here they are, with what they ask you to do.
Agents and Workflows (17%)
- Architecture and Runtime Selection: decide between a workflow and an agent from requirements, error tolerance and audit needs, choose the orchestrator and subagent split, and pick a self-hosted harness or a managed, sandboxed agent runtime.
- Agent SDK Construction and Hooks: build and correct Agent SDK agent definitions, custom loops and hooks so that required actions run deterministically.
- Agent Pattern Diagnosis: read a trace of a misbehaving agent and name the missing pattern, such as loop termination, subagent isolation, memory or context management.
Applications and Integration (20%)
- Messages API Mechanics: message structure, tool use, streaming, vision, thinking, prompt caching, structured output and stop reasons, plus realtime versus batch.
- Application Boundary and Surface Behavior: how Claude reads instructions on different surfaces, what crosses the application boundary, and whether schema and session decisions meet the requirements.
- Configuration and SDLC Integration: CLAUDE.md files, settings.json, model version pinning, prompt versioning, plugin dependencies, CI integration and how a configuration change affects each environment.
- Integration Resilience: find the defect in failing integration code, including mishandled streaming events, retry logic that does not tell rate limits, overloads, invalid requests and timeouts apart, and faulty multi-format ingestion through the Files API.
Claude Code (8%)
- Claude Code Operation: CLI, desktop, IDE, web and mobile surfaces, sessions, the CLAUDE.md hierarchy, settings.json, permission modes, headless and streaming modes, and built-in commands.
- Claude Code Customization and Extension: choose a Skill, plugin, Rules file, hook, subagent or slash command for a team workflow, then author and package it.
Eval, Testing, and Debugging (10%)
- Evaluation Design: build evaluations with LLM-as-judge harnesses calibrated against human labels, golden sets, regression tests and behavioral checks, and choose defensible metrics.
- Failure-Mode Diagnosis: identify hallucination, refusal, under-acting, context loss, format drift or inconsistent output from a trace, decide whether it starts in the integration layer or the model output, and fix the cause.
Model Selection and Optimization (12%)
- Generation Mechanics and Model Options: tokens, context windows, sampling and non-determinism, adaptive thinking and effort levels.
- Model Tier Selection and Release Planning: choose a tier across quality, latency and cost, and plan for behavior changes across model releases with regression evaluations.
- Cost Diagnosis and Optimization: find the dominant cost driver in usage data and fix it with prompt caching (single, layered and rolling breakpoints), batch processing or token budgets, with the expected impact.
Prompt and Context Engineering (15%)
- Context Management for Long-Running Sessions: compaction, tool-output pruning, subagent isolation, and when to start a fresh session.
- Prompt Revision: close a gap that regression evaluations exposed, using instruction clarity, system versus user placement, examples, output constraints and input sanitization.
- Structured Output and Defensive Parsing: strict schemas, response validation and defensive parsing of free-text fallbacks.
- Grounding and Source Citation: answers grounded in supplied documents, with citations and a "not found" answer when the documents do not contain it, and full-context versus retrieval by corpus size, cost and freshness.
Security and Safety (8%)
- Prompt Injection and Jailbreak Defense: layered mitigations for an agent that reads untrusted content.
- Deterministic Guardrails and Least Privilege: guardrail layers, least privilege and hooks that keep destructive actions from running even under adversarial instructions.
- API Key and Credential Handling: server-side proxying, workspace-scoped versus admin keys, environment separation, rotation and access monitoring.
Tools and MCP (10%)
- Tool Design and Response Shaping: schemas, descriptions that drive the right call, minimal actionable results, error handling, where the tool runs and approval patterns.
- MCP Server Development and Integration: resources, tools and prompts, stdio or HTTP transport, and authentication.
- Capability Mechanism Selection: built-in tools, custom tools, Skills or MCP servers, chosen by where the capability lives, who maintains it and how Claude finds it.
Three questions from the CCDV-F practice tests
A fintech triage endpoint must return exactly one of three labels (approve, reject, review) with no surrounding prose, because a downstream rules engine string-matches the value. The model keeps appending a short justification sentence after the label. Which technique most directly enforces the constrained output?
[object Object]
Correct. Constraining the response with a schema whose field is an enum of the permitted labels is the documented way to guarantee the output shape instead of merely encouraging it.
B. A tight cap truncates whatever the model happened to generate; if it starts with prose the label gets cut, so this trades extra text for corrupted output.
C. Stop sequences end generation when matched and the matched text is excluded from the response, so the model could still produce prose first and the label itself would be cut off.
D. Low temperature reduces variation between runs but does not constrain the format; a model inclined to justify its label will justify it consistently.
A Django team is building a support assistant on the Messages API. Standing behavior rules (you are a support assistant; cite the KB article id for every claim) currently sit in the first user message, and the model drifts off the rules after six or seven turns. Per documented request structure, where should these standing rules live?
[object Object]
Correct. The system parameter is the documented place for role and behavior instructions that should govern every turn, so they persist instead of fading like content buried in one user message.
A. The metadata field carries identifiers like an external user id for abuse detection; the model does not read it, so rules placed there have no effect on behavior.
B. A fabricated assistant turn reads as prior model output, not operator guidance, and the messages array must begin with a user turn.
C. Re-injecting the block every turn bloats the context and is a workaround; the documented home for persistent behavior guidance is the system parameter.
An e-commerce platform tags 40,000 incoming support tickets per day into six fixed categories. In offline tests Claude Haiku already matches the accuracy of the larger tiers on this narrow task, and per-ticket latency and cost dominate the budget at this volume. Which model choice fits the workload?
[object Object]
Correct. A simple, high-volume task that is already accurate on the fastest tier is the canonical fit for Haiku, minimizing per-ticket cost and latency with no measured quality loss.
A. Opus adds cost and latency the task does not need; a fast tier already matches its accuracy on this narrow classification, so the headroom is paid for but unused.
B. Sonnet is a sensible middle tier for mixed workloads, but here the measurements already show the fastest tier is sufficient, so the extra capability is unneeded spend at 40,000 tickets a day.
C. An ensemble more than doubles cost and latency per ticket, which contradicts the budget constraint when a single small model already meets the accuracy bar.
Full CCDV-F set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsWhat disappears, and what is new
The current guide's skill list mixed software fundamentals with Claude skills. Several of those general skills do not return: Understanding Requirements, Systems Life Cycle, Software Engineering Foundations, LLM Fundamentals and Technical Fundamentals are gone as separate skills. Their Claude-specific parts survive inside the new skills, such as generation mechanics, configuration and SDLC integration, and integration resilience.
The new skills also name topics the current guide does not: managed agent runtimes (the guide's audience section names Claude Managed Agents), Files API ingestion, Claude Code on every surface and its extension mechanisms, calibrated LLM judges, grounding with a "not found" answer, and credential handling with workspace-scoped keys. And the verbs change. Many skills now start with "analyzing" a trace, a failing piece of code or cost data, so expect more questions that hand you evidence and ask what is wrong.
Should you test before or after December 7?
Anthropic's advice is simple: if you test on December 7 or later, prepare from the updated guide.
- Nearly ready on the current guide? Book a date before December 7 and sit the version you studied for. The current exam leans on API integration (33.1%) and model selection (16.8%).
- Starting now, or more than three weeks from ready? Study from the updated guide and book on or after December 7. Put real time into evaluation, Claude Code and prompt and context engineering, which together carry a third of the new exam.
- Strong on Claude Code and evals, lighter on the API? The new weights favor you, so the December version is the better fit.
How to prepare for the new version
- Read the updated exam guide end to end, including its revised sample questions.
- Turn each skill above into a hands-on check: write one evaluation with a calibrated judge, package one Claude Code extension, break one integration on purpose and read the error class.
- Use the CCDV-F complete guide and the exam domains breakdown for depth; both are being revised to the new guide.
- Practice under time pressure with the CCDV-F practice tests. Six follow the current guide for exams until December 6, and 70-question practice tests on the December 7 blueprint are being added next to them; the first went live on November 19, 2026.
If you are also looking past Foundations, the professional level opened on November 17: see the Developer - Professional exam guide.
Frequently asked questions
Sources: