Is this CCDV-F practice test really free?

Yes. The 25-question timed diagnostic is free with a free account, and you get your score and domain breakdown at the end. No card is required.

Are these real Claude Certified Developer exam questions?

No. Real items are confidential under Anthropic's candidate agreement. Every question here is original, written by PrepClubs from Anthropic's published exam guide and documentation.

How many questions are on the real CCDV-F exam?

53 questions in 120 minutes, multiple choice and multiple response, with a scaled passing score of 720 out of 1,000.

What do I get if I upgrade?

Ten full-length timed exams of 53 questions (530 in total) at the official domain weights, with an explanation and an Anthropic docs reference for every answer. It is $69 one time, not a subscription, with 30-day access and a Money-back Guarantee*.

Do I need to create an account?

Yes, a free one. The sample questions on this page are open to anyone.

Practice Hub

Free Claude Certified Developer (CCDV-F) Practice Questions

Not sure you are ready for the Claude Certified Developer - Foundations exam? Try the three questions below, then take the free, timed 25-question diagnostic. It samples all eight domains at Anthropic's official blueprint weights and runs at the real exam's pace, about 2 minutes 15 seconds a question, so your result shows where you actually stand.

By PrepClubs Editorial Team, updated September 25, 2026

Questions

Time Limit

120 min

Difficulty

High

Access

30 days

Full access

Start prepping

What the free CCDV-F diagnostic includes

Twenty-five original questions in 57 minutes, drawn from the same bank as the ten full exams. Applications and Integration gets the largest share because it is a third of the real exam, and every other domain appears at its published weight.

You see your score and a domain-by-domain breakdown the moment you finish, the same shape as Anthropic's own score report. That tells you which two domains to study first.

All eight domains

Agents and Workflows, Applications and Integration, Claude Code, Eval and Debugging, Model Selection, Prompt and Context Engineering, Security and Safety, Tools and MCPs.

The exam's question shapes

Scenario items, multiple-response items that say how many to select, and code-reading items with real request bodies and config files.

Written from Anthropic's docs

Original PrepClubs questions built from the official exam guide and documentation. Not real exam items and not dumps.

Instant domain breakdown

Percent correct by domain, so your study time goes where the points are.

Money-back Guarantee* on the full bank

Upgrade to 10 full-length exams, backed by our Money-back Guarantee*.

Three free CCDV-F practice questions

These three are original items held back from every paid exam. Pick an answer, then read why each option is right or wrong.

Sample 1: Applications and Integration

A weather tool's upstream API times out while Claude is mid-task. The developer wants Claude to tell the user the data is temporarily unavailable, or try another approach, without breaking the conversation. How should the harness answer the pending tool_use block?

A.Return a tool_result with empty content, which Claude reads as a failure by convention.
B.Skip the tool_result and send a plain user text message that explains the outage instead.
C.Return a tool_result with is_error: true and text like "Upstream timeout; retry in 60s."
D.Delete the tool_use block from the assistant message and resend the history without it.

Answer and walkthrough

C. When a client tool fails, return a tool_result for its tool_use_id with the error text in content and is_error: true. Claude then incorporates the failure into its reply or tries another route. The docs recommend instructive messages that say what went wrong and what to try next. Why the others are wrong: Every tool_use must be answered by a tool_result in the next user message; a text-only reply triggers a 400 error. Editing Claude's earlier turn hides what happened and leaves Claude unable to reason about the failure. An empty result is a valid, successful result, so Claude would not know the call failed. Source: Anthropic docs, Handle tool calls, Handling errors with is_error.

Sample 2: Tools and MCPs

An operations agent returns two tool_use blocks in one turn: create_ticket and then assign_ticket, which needs the ticket to exist. Because both have side effects, the harness runs them sequentially. create_ticket fails with a validation error, so the harness never runs assign_ticket. How should the harness reply?

A.Send is_error results for both, saying assign_ticket was skipped due to the failure
B.Delete the assign_ticket block from the stored assistant message, then send the create_ticket error
C.Send an is_error result for create_ticket and omit assign_ticket, since it never ran
D.Run assign_ticket anyway so that each tool_use block gets a genuine execution result

Answer and walkthrough

A. Every tool_use block needs a matching tool_result in the next user message, even when the harness decides not to execute it. For a call that was skipped because an earlier call in a sequential batch failed, the docs say to return a tool_result with is_error: true and a brief explanation, so Claude understands what happened and can reissue the call. Why the others are wrong: Omitting the result for assign_ticket leaves a tool_use id without a result, which the API rejects. Editing Claude's stored assistant turn corrupts the conversation history rather than reporting what actually happened. Running a side-effecting call whose prerequisite failed produces a pointless or harmful action just to have a result to report. Source: Anthropic docs, Parallel tool use, Execution semantics.

Sample 3: Model Selection and Optimization

On the 23rd of the month, every request starts returning HTTP 429 rate_limit_error. The responses have no retry-after header, error.details.error_code is enforced_spend_limit_reached, and the SDK's automatic retries all fail. What is the correct response?

A.Pause until the anthropic-ratelimit-requests-reset time, then resume traffic.
B.Shift traffic to another model, since rate limits apply separately to each model.
C.Raise max_retries and backoff so calls succeed once the per-minute bucket refills.
D.Treat it as the tier's monthly spend cap: alert, and request a higher limit or tier.

Answer and walkthrough

D. A spend-cap 429 uses the same rate_limit_error type as a rate limit but has no retry-after header and carries enforced_spend_limit_reached. Usage stays paused until the next month unless a higher limit or tier is granted, so retries are wasted. Alert on this code and escalate. Why the others are wrong: Retries keep failing until access resumes, because this is a monthly cap rather than a per-minute bucket. The spend cap is organization-wide, so switching models does not restore access. The reset header describes rate-limit replenishment, which is not what is blocking these requests. Source: Anthropic docs, Rate limits, Reaching your spend cap.

What the real CCDV-F exam is like

The exam has 53 items in 120 minutes, proctored by Pearson VUE online or at a test center. Items are multiple choice or multiple response, and each multiple-response item states how many options to select. You pass with a scaled score of 720 on a 100 to 1,000 scale; the report also shows your percent correct in each domain.

The questions test applied judgment. Most describe a system, a constraint and a symptom, then ask which approach fits: the Message Batches API or real-time calls, a hook or a prompt rule, an MCP server or a custom tool. Knowing what each feature does is not enough; you have to know when it is the wrong choice.

Applications and Integration (33.1 percent), Model Selection and Optimization (16.8 percent) and Agents and Workflows (14.7 percent) make up almost two thirds of the exam. A gap in any of those three costs the most points.