Anthropic certification practice exams

Measured on the banks published here, the “always pick the longest option” strategy scores: Associate Foundations 0.0%; Developer Foundations 26.3%; Architect Foundations 23.6%; Architect Professional 1.6%. Random guessing on four options scores about 25%. On Associate Foundations, Developer Foundations, Architect Foundations and Architect Professional the writing itself gives nothing away.

How these work

Each exam runs like the real thing rather than like a flashcard deck. Correct answers, rationale and the notes explaining why each wrong option fails are hidden for the whole attempt. You answer, move freely between questions, change your mind, then submit. Only then do you get a score, a per-domain breakdown, and a full review of every question.

Question order and option order are reshuffled on every attempt, so a remembered position such as “the answer was C” is worth nothing. Multi-select questions are scored all-or-nothing, with no partial credit, which is how the real exams score them.

Everything runs in your browser from a single static page. No account, no server, no analytics, nothing uploaded.

Available right now: Associate Foundations, Developer Foundations, Architect Foundations, Architect Professional, 669 questions in total. The other certifications on this page are listed so the set is complete, but they are not clickable yet and the tile says why.

How the question banks are built

Every bank is validated against a published schema before it ships. The validator does more than check structure: it measures whether a candidate with zero knowledge could pass by exploiting the writing.

That check exists because an early version of one bank failed it badly. The strategy “always pick the longest option” scored 86% on it. Correct answers were on average 2.05× longer than the distractors, and two dozen distractors were obvious throwaways. The exam was testing reading length, not judgement. After a rewrite, the same exploit scored 1.6%.

So each bank is checked for: the longest-option exploit, the shortest-option exploit, the most-punctuation exploit, the ratio of correct-answer length to distractor length, and an even spread of correct answers across positions. Every wrong option has to be a practice a competent engineer could defend in some other context but which fails for the specific constraint in that question’s scenario.