The seven most common mistakes in Timo's Claude certification practice exams are about keeping customers' data private, measuring whether a Claude system works, checking whether Claude's answer is correct, connecting Claude to other systems, getting output in a fixed format, keeping track of where a finding came from, and questions that ask for two answers. Across 12,585 answers, 17.3% were wrong. Questions that ask for two answers were missed more often (23.3% of answers wrong) than questions with one answer (15.8%).
What the data covers
Between 22 September and 15 November 2026, people started 9,415 Claude certification practice tests on Timo by Amotion AI. There are two kinds of test: free 20-question tests on timolabs.dev, and full-length practice exams of 53 to 63 questions in the Timo portal, which 265 people started. The median score is for tests with all questions answered: 75% for CCAO-F and 80% for the other three exams on the free tests.
| Practice test | Started | Scored | All questions answered | Median score |
|---|---|---|---|---|
| Free 20-question tests | 8,493 | 6,005 | 3,589 | 75% to 80% |
| Full-length practice exams | 922 | 900 | 483 | 86.7% |
A test counts as scored when it reached a result, either because the person submitted it or because the time limit ended it. The free tests need no sign-in, so they count tests, not people. They record each test's score but not the answer to each question. The full-length practice exams record every answer, so the mistake analysis below uses them. It uses each person's first fully answered attempt on each exam. Where a person never answered every question, it uses their first attempt with answers and counts only the questions they answered. That gives 294 practice exams from 237 people, with 12,585 answers. Each answer is one person's response to one question, and a two-answer question counts as correct only when both choices are right.
How the exam questions work
Most questions describe a company that is building something with Claude, Anthropic's AI model. For example, a company might add a Claude assistant to its website, or use Claude to write reports. The company's own software, called the application, sends requests to Claude and passes Claude's answers back to its users. The question asks which decision the company should make. Several options usually sound sensible, and only one fits what the scenario needs.
Timo took the 12 questions that people missed most on each exam, 48 in all. For 39 of these 48 questions, at least half of the people who answered incorrectly chose the same wrong option. The sections below explain the seven most common mistakes behind those choices: keeping customers' data private, measuring whether a Claude system works, checking whether Claude's answer is correct, connecting Claude to other systems, getting output in a fixed format, keeping track of where a finding came from, and questions that ask for two answers. Each one uses a question from the free practice tests.
1. Keeping customers' data private
A company runs a reporting application for many client companies, and each user signs in to it. Claude fetches report data by calling a tool, which is a function the application runs for Claude, such as a data lookup. An API key is a secret code that gives access to data. The application uses one administrator API key that opens every client company's records. It puts this key in the system prompt, the standing instructions it sends to Claude with every request, and Claude uses the key in every tool call. Signing in tells the application who the user is, but nothing checks which company's records each tool call fetches. So a user from one company could ask Claude about another company, and Claude could fetch that data.
So an exam question can ask: which change stops a user from seeing another company's data?
On the most-missed questions about this topic, 19% of test takers chose an answer that relies on Claude's instructions or keeps one key that opens everything. Examples are adding a rule to the system prompt, or moving the key into a secret store, which is a protected place where the application keeps its keys. That is incorrect. On those questions, which came up on all four exams, 30% of answers were wrong.
The correct answer is that the application first checks which company the signed-in user belongs to, then runs each tool call with a short-lived key that only opens that company's records. This is because adding an instruction does not ensure Claude will follow it, and a key moved to a secret store is hidden but still opens every company's data.
To apply this: a rule that must never be broken is built into the application's code. Before any data reaches Claude, the code checks what the signed-in user is allowed to see, and each key only opens what that user may see.
A related case: a company's Claude assistant should reply in a friendly tone, and an occasional reply in a slightly wrong tone does little harm. Because instructions are not guaranteed, the team considers adding code that checks the tone of every reply before it is sent. So an exam question can ask: should the team add that check? Adding it is incorrect, because the check slows down every request to prevent a mistake that costs little. The correct answer is to keep the tone rule in the prompt and track how often Claude misses it, because an instruction is enough when an occasional miss does little harm.
Study topics: secrets and key management, Claude Code hooks and integration security for architects.
2. Measuring whether a Claude system works
A university plans a Claude assistant that answers students' course-planning questions before they meet an advisor. The aim is to free up advisors' time, but students with complex or sensitive issues must still reach an advisor.
So an exam question can ask: which goal should the university set for the assistant?
On the most-missed questions about this topic, 28% of test takers chose a single overall number, such as "raise the share of conversations resolved without an advisor by 30%". That is incorrect. On those questions, which came up on CCAO-F, CCAR-F and CCAR-P, 41% of answers were wrong.
The correct answer is "cut advisors' preparation time by 30%, as long as students who need an advisor are still referred to one and student satisfaction stays above a set minimum". This is because the share of conversations resolved without an advisor rewards keeping students away from advisors, even the ones who need one.
To apply this: one overall number does not show where a system fails. It can look good while one type of case gets poor results, so results are broken down by type of case. Any target for saving time or cost is paired with a minimum level of quality, so that savings cannot hide worse answers.
Study topic: evaluation, testing and optimization.
3. Checking whether Claude's answer is correct
A company uses Claude to draft a quarterly sales report for its executives, and a reviewer checks it before it goes out. The report cites section 7.3 of a policy that only has six sections, and the summary says total revenue was $4.2M while the table says $4.6M. Because Claude wrote the draft, asking Claude to check it can seem the quickest fix.
So an exam question can ask: what should the reviewer do about these two problems?
On the most-missed questions about this topic, 22% of test takers chose a check that relies on Claude, such as asking Claude how confident it is or asking it to check its own work. That is incorrect. On those questions, which came up on all four exams, 33% of answers were wrong.
The correct answer is to open the policy, which shows that section 7.3 does not exist, and to recalculate revenue from the original sales records. This is because Claude can be wrong and still sound sure, and asking it again can repeat the same mistake.
To apply this: a check goes back to the original source, meaning the raw data or the document itself. How much a person reviews depends on how much harm a wrong answer would cause, not on how confident Claude sounds.
Study topics: output evaluation and validation and human review and confidence.
4. Connecting Claude to other systems: tools, streaming and caching
A ferry company's Claude assistant answers customers' questions about sailings. Claude cannot look up the schedule by itself, so the application gives it a tool for looking up sailing times. Claude asks for a lookup, the application runs it and sends back the result, and Claude then answers or asks for another lookup. In some chats, the assistant stops right after the first lookup and shows the customer nothing. In others, Claude asks again for a lookup that has already run.
So an exam question can ask: how should the application run this exchange so that every customer gets an answer?
On the most-missed tool questions, 22% of test takers chose their question's most common wrong answer, such as stopping as soon as the tool returns its first result. That is incorrect. On those questions, which came up on CCDV-F and CCAR-F, 33% of answers were wrong.
The correct answer is to send each result back to Claude together with the whole conversation, and to keep going until Claude signals that its answer is finished. This is because Claude may need several lookups, and the conversation shows Claude which lookups have already run.
In API terms, for developers: the application talks to Claude through Anthropic's API, the interface that software uses to send requests to Claude and receive its replies. Each reply says why Claude stopped, in a field called stop_reason. A value of tool_use means Claude is asking for a tool. The application returns the tool's data as a tool_result, labelled with a tool_use_id that matches Claude's request. A value of end_turn means Claude has finished. A value of max_tokens means the reply hit its length limit and is not finished.
Two related cases come up in the same questions:
- Streaming. Some chat assistants show Claude's reply while it is being written. With streaming, the reply arrives in small pieces, and the application also saves each reply to the chat history. So an exam question can ask: when should the application save the reply? Saving each piece as if it were the full reply is incorrect. The correct answer is to collect the pieces and save the message when the API sends a
message_stopevent, then checkstop_reasonto see whether Claude finished (end_turn), asked for a tool (tool_use) or hit the length limit (max_tokens). This is because each piece holds only part of the message, and a complete message is not always a finished answer. - Prompt caching. An assistant sends the same instructions and list of tools with every request, followed by the user's latest message. Prompt caching saves cost by reusing the work Claude already did on an identical start of a prompt. Because the latest message matters most, putting it first can seem sensible. So an exam question can ask: how should the team arrange the prompt? Putting the latest message or other changing details at the start is incorrect. The correct answer is to keep the repeated part identical at the start and put changing details after it, because any change in the start stops the saved work from being reused.
Study topics: Claude API mechanics and cost and token management.
5. Getting output in a fixed format
A textile mill asks Claude to read defect reports and return the type of defect and how serious it is as JSON, a data format with named fields that another program reads automatically. Even with two pages of instructions, the field names and structure keep changing from report to report.
So an exam question can ask: what should the team change in its prompt?
On the most-missed questions about this topic, 24% of test takers chose more instructions, such as adding "follow the format exactly" at the end. That is incorrect. On those questions, which came up on CCAO-F, CCAR-F and CCAR-P, 36% of answers were wrong.
The correct answer is to show Claude two to four reviewed reports, each with its correct JSON. This is because more instructions repeat the requirement without showing what the finished output looks like, while the examples show the exact structure and how seriousness is judged.
To apply this: if the structure must never vary, the API's structured outputs feature enforces it with a JSON schema, which is the list of required fields and the type of each one. The examples still show Claude how to judge each report.
A related case: a company's assistant needs several tools for different tasks, but the tools have short, similar descriptions, and Claude keeps choosing the wrong one. The team considers adding rules to the prompt about when to use each tool, or blocking some tools. So an exam question can ask: how should the team fix the wrong tool choices? Adding rules or blocking tools is incorrect. The correct answer is to rewrite each tool's description to say what it does, what it returns and when to use it, because Claude decides which tool to use mainly from those descriptions. More rules leave the confusing descriptions in place, and blocking a tool stops the assistant from doing the tasks that need it.
Study topics: few-shot prompting, prompting and task execution and tool interface design.
6. Keeping track of where a finding came from
A company's product team uses Claude to analyse 400 exit-survey comments, which are answers from users who left, and support tickets, which are users' requests for help. The aim is to decide which steps of the sign-up process to redesign. Claude finds what customers complained about and suggests likely causes, and other teams will act on the findings. One ranked list of everything can seem the easiest format for those teams.
So an exam question can ask: how should the team present the findings?
On the most-missed questions about this topic, 28% of test takers chose one ranked list that merges what customers said with Claude's guesses about the causes. That is incorrect. On those questions, which came up on CCAO-F, CCDV-F and CCAR-F, 45% of answers were wrong, the highest of these topics.
The correct answer keeps what customers reported separate from Claude's explanations, with each finding linked to the comments it came from. This is because a mixed list does not show which points customers made and which Claude inferred, so nobody can check it before acting on it.
To apply this: every figure also keeps its source and the date its data was collected, so readers can tell how old it is.
A related case: a company uses an AI agent to handle customer cases. An agent is an AI system that chooses its own next steps as it works through a task. When a difficult case needs a person, forwarding the whole conversation can seem the safest way to make sure no detail is lost. So an exam question can ask: what should the agent hand over to the person? Passing on the whole conversation is incorrect. The correct answer is a short summary with the case reference, what was found, why a person is needed and a suggested next step, because a person cannot quickly find what matters in a long conversation.
Study topics: information provenance and workflow enforcement and handoff.
7. Questions that ask for two answers
Some questions ask you to pick two correct options instead of one. For example, a company uses a Claude assistant to help customers troubleshoot problems, and its team runs an A/B test, which compares a new troubleshooting prompt with the current one. Results can differ by chance even when two prompts are equally good, so before starting, the team fixed the number of test cases and chose a significance test. A significance test is a statistical check of whether a difference between the two prompts is too large to be explained by chance alone. Each extra tool call adds time and cost, so the team also set a limit on extra tool calls. After 40 cases, before reaching the planned number, the new prompt is slightly ahead on quality but makes more tool calls.
So an exam question can ask: which two actions should the team take now?
A common wrong pair is continuing the test, which is correct, together with changing the new prompt midway so it makes fewer tool calls. That pair is incorrect. On the most-missed two-answer questions, 94% of wrong answers paired one correct option with one that sounds sensible but does not fit the scenario. Overall, 23% of answers to two-answer questions were wrong, against 16% for one-answer questions.
The first correct answer is to continue to the planned number of cases and then apply the chosen significance test. The second is to keep the new prompt unchanged until the end, then check its tool calls against the limit. This is because the lead after 40 cases could be due to chance, and changing the prompt midway means the test no longer compares the same prompt.
To apply this: check each option on its own against what the question asks for. An option that is good practice in general is still wrong if the scenario does not need it.
On the architect exams: choosing an agent when the steps are fixed
An insurer wants Claude to help process policy changes. Every change goes through the same four steps: extract the requested changes, check the policy fields, draft the new wording, and send exceptions to an underwriter, the insurer's specialist who reviews risk. Because several steps need Claude's judgment, an agent can seem the natural choice. The other option is a coded workflow, where the application's code runs the steps in a fixed order and calls Claude only at the steps that need it.
So an exam question can ask: should the insurer build this with an agent or with a coded workflow?
On the architect question that tests this, 47% of test takers chose a design built on one or more agents. That is incorrect.
The correct answer is a coded workflow that runs the four steps in order and calls Claude at the steps that need it. This is because the steps never change, so an agent's freedom to choose its next step only adds unpredictability.
Study topics: task decomposition and solution design and architecture.
Where candidates lose the most marks
Each exam is divided into domains, the topic areas listed in Anthropic's exam guide. The table compares the share of answers that were wrong in each domain.
| Exam | Highest share of wrong answers | Lowest share of wrong answers |
|---|---|---|
| CCAO‑F | Workflow Integration and Solution Design (18.6%); Output Evaluation and Validation (18.3%) | Configuration and Knowledge Management (8.9%) |
| CCDV‑F | Security and Safety (26.2%); Model Selection and Optimization (16.4%) | Agents and Workflows (9.1%) |
| CCAR‑F | Prompt Engineering and Structured Output (23.8%); Agentic Architecture and Orchestration (21.5%) | Tool Design and MCP Integration (15.6%) |
| CCAR‑P | Evaluation, Testing and Optimization (25.2%); Integration (23.6%); Governance, Safety and Risk Management (23.4%) | Developer Productivity and Operational Enablement (15.1%) |
The lowest column leaves out two small CCDV-F domains, Claude Code and Eval, Testing and Debugging, because each had fewer than 70 answers. The CCDV-F Security and Safety figure comes from only 130 answers, so the true rate for all candidates could be noticeably higher or lower.
Scores and time by exam
| Exam | People | Answers wrong | Median score | Median time |
|---|---|---|---|---|
| Claude Certified Associate Foundations (CCAO-F) | 112 | 14.1% | 88.3% | 62 minutes |
| Claude Certified Developer Foundations (CCDV-F) | 45 | 14.2% | 88.7% | 65 minutes |
| Claude Certified Architect Foundations (CCAR-F) | 84 | 19.6% | 82.5% | 82 minutes |
| Claude Certified Architect Professional (CCAR-P) | 53 | 21.5% | 82.6% | 91 minutes |
The median score and median time are for fully answered practice exams only: 69 people on CCAO-F, 30 on CCDV-F, 62 on CCAR-F and 32 on CCAR-P. These are practice scores. The official exams report a scaled score from 100 to 1,000, with 720 needed to pass, and a practice percentage does not convert to that scale. Different people take each exam, so the table does not show that one exam is harder than another.
Questions with two answers are missed more often
Some questions ask candidates to choose two options. These questions had more wrong answers on three of the four exams.
| Exam | Two-answer questions wrong | One-answer questions wrong |
|---|---|---|
| CCAO‑F | 20.0% | 12.6% |
| CCDV‑F | 11.3% | 14.9% |
| CCAR‑F | 30.3% | 17.0% |
| CCAR‑P | 26.8% | 20.2% |
| All four exams | 23.3% | 15.8% |
Among the 48 most-missed questions, 17 ask for two answers. On those 17 questions, 318 of the 337 wrong answers (94%) included one correct option. Candidates usually found one right answer and then added a second option that sounds reasonable but does not fit the scenario.
How to use these findings in your preparation
After each practice test, name the topic behind every wrong answer, not only the correct option. If you missed a question because you trusted Claude's own check or added a rule to the prompt, look for the same pattern in the other questions you missed. On questions that ask for two answers, test each option on its own against what the scenario asks for.
- Take a free 20-question practice test for your exam and note your weakest domain.
- For every wrong answer, name the topic you misread, not only the correct letter.
- When an option says "tell Claude" or "ask Claude", check whether the scenario needs a check in the application's code or a check against the original source instead.
- When you practise again, use questions you have not seen before, so the score reflects understanding rather than memory.
Free practice tests: CCAO-F, CCDV-F, CCAR-F and CCAR-P. Each has 20 questions, scores you by domain and needs no sign-up.
How the data was collected
- Free tests: The counts come from Timo's PostHog analytics for timolabs.dev and cover 22 September 2026 to 15 November 2026 at 12:00 Dubai time. A test counts as fully answered when all 20 questions were answered at submission.
- Full-length practice exams: The answers come from the Timo portal's records up to the same time and were read without any changes. One internal test account was excluded. Scores recalculated from the stored answers matched the portal's recorded scores in every case.
- Most-missed questions and topics: Timo took the 12 most-missed questions on each exam, each answered by at least 10 people, 48 in all. It grouped them by the concept each one tests. The topic figures above are calculated over the most-missed questions in each group, not over every question on that topic. One question was left out of the groups while its answer is checked against the latest Claude documentation.
- Limits: These are Timo practice tests, not Anthropic's official exams, and the scores are not official pass rates. People who take full-length practice exams in the Timo portal are preparing with Timo, so they are not a random sample of all candidates. The CCDV-F and CCAR-P groups are small (45 and 53 people). Most people start with practice exam 1 of the five versions for each exam, so most full-length results come from that version.
Earlier practice data, including scores by country, is on the Claude certification practice exam data page.
Common questions
What is the most common mistake on the Claude certification exams?
In Timo's practice data, the mistake behind the most of the 48 most-missed questions was relying on an instruction to Claude, or on one key that opens everything, where the scenario needs a check in the application's code. It appeared in 9 of the 48 questions, on all four exams. Measuring success with a single overall number appeared in 8.
Are questions with two answers harder?
They were missed more often on three of the four exams. Across all four, 23.3% of answers to two-answer questions were wrong, against 15.8% for one-answer questions. On the 17 two-answer questions among the 48 most missed, 318 of the 337 wrong answers included one correct option.
How hard is the Claude certification exam?
The official exams use scenario-based questions, and Timo does not have official exam results. In Timo's full-length practice exams, the median candidate who answered every question scored 83% to 89%, depending on the exam, and 17.3% of all answers were wrong.
What is the passing score for the Claude certification exams?
Each official exam reports a scaled score from 100 to 1,000, and 720 is needed to pass. A practice percentage does not convert directly to this scale.
How long does a full-length practice exam take?
The median time in Timo's full-length practice exams was 62 minutes for CCAO-F, 65 for CCDV-F, 82 for CCAR-F and 91 for CCAR-P.
Are these official pass rates?
No. These figures come from Timo practice tests. They do not include results from Anthropic's official exams.
Find your weak domains
Take a free 20-question practice test for your exam. It scores you by domain and needs no sign-up.
By Amotion AI