How hard is the Claude Certified Architect (CCAF) exam?

The format in numbers

These figures come from Anthropic's exam guide (version 1.0, July 2026):

The pass mark is a scaled score. 720 does not mean 72% of the questions right, and nothing in the figures above turns it into a count of correct answers.

Why it is demanding

Four things make it harder than its length suggests.

• It tests judgment, not recall. Knowing what a feature does is only the starting point. The questions ask what you would do with it in a specific situation. In the Multi-agent research system scenario, the hard part may not be how sub-agents work but whether the task needs them at all, or whether one agent with the right tools would do.
• You cannot skip a scenario. Each sitting draws 4 of the 6 scenarios at random, so you do not know in advance which ones you will get. Two of the six centre on Claude Code: Code generation with Claude Code, and Claude Code for continuous integration. If you have built agents with the API but never configured Claude Code, the draw can expose that gap.
• Every domain counts. The score is split across five domains. The largest, Agentic architecture and orchestration, is 27%. The smallest, Context management and reliability, is still 15%, so none can be left out. The full table is in our guide to what the exam tests.
• Two minutes goes quickly. That average includes reading the scenario. Multiple-response questions take longer, because each option has to be judged on its own rather than against the others.

A typical judgment call looks like this. Claude Code runs in a CI pipeline to review pull requests, and it needs credentials. Writing the key into the workflow file or the prompt would work on the first day. The question is which setup keeps the secret out of the repository, the logs and the model's context. Each option describes something a real team might do, which is why reading carefully matters more than speed.

What first sittings of our mock show

Our timed mock has the same length and time limit as the real exam: 60 questions in 120 minutes. Between 1 August and 14 November 2026, 624 people completed a first sitting on CyberSkill Practice. We counted a sitting when at least 54 of the 60 questions were answered.

• The median first score was 75%.
• 57.7% scored 72% or more, which is the pass line our mock uses. About four in ten first sittings fell short of it.

This is not a pass rate for the real exam. Anthropic's 720 is a scaled score, so our percentage is not the same measure. Our mock covers four of the six scenarios, it is not the real exam, and its scores do not predict the real result. What it does show is that a large share of people who practised did not clear our line the first time through.

Repeat scores and scenario scores

People who came back scored much higher: the median across 237 later sittings was 91.7%. We do not read that as proof of learning. Our bank has 80 questions, so a repeat sitting shows many of the same questions again, and remembering an answer is not the same as reasoning your way to it. A repeat score overstates readiness.

We also looked at the share of questions answered correctly in each practice scenario, across 925 full sittings:

The gap from lowest to highest is about three points. That is too small to name one scenario as the hard one, and it is no reason to skip any of them.

Try it

The quickest way to judge the difficulty for yourself is to read a few questions. The sample questions come with their answers and need no account. When you want all 60 against the clock, start the timed mock from the exam page. For a study plan, read our guide on how to pass the exam.

CyberSkill Practice is free and unofficial, and it is not affiliated with Anthropic.