Five AI models took my practice exam and all scored 98% or more. A developer who built ClaudeArena, a free practice site for Anthropic's Claude Certified Architect exams, tested five AI models from four makers on the 63-question Professional exam on Hard mode, and all scored 98% or 100% well within the 120-minute limit. Qwen and DeepSeek missed the same question with the same wrong answer, and a separate tool-free Luna run also chose that answer, which the developer verified against Anthropic's documentation. A Qwen run also exposed an accessibility bug: the page highlighted selected answers visually but never announced them structurally, so the site now announces chosen answers and fails a build check if the highlight and announcement diverge. This is part 4 about ClaudeArena https://claudearena.dev , the free practice site I built for Anthropic's Claude Certified Architect exams CCAR-F and CCAR-P . Last time it was about what real usage taught me. This time it's about a question I had been avoiding. Two callnames on the leaderboard had fast, high scores. Good players, or an AI doing the work? I couldn't tell from the numbers. So I stopped guessing and tested what an AI actually scores. I gave an AI agent a browser, signed in to a test account myself, and gave it one instruction: take the full Professional exam on Hard, answer from your own knowledge, don't look anything up. That's 63 questions with a 120-minute limit. Then I repeated it with other models, including cheap ones on their lowest settings, to see the other side of the coin. Five models from four different makers, and every one scored 98% or 100%. Luna was the fastest. The others took longer, and every one finished well inside the time limit. To be fair about it: Fable 5.2 helped me build and review the question bank, so its score was no surprise. None of the others had seen the questions. DeepSeek needed one nudge from me to keep going, and an attempt with Grok's free tier stopped at question 3 when the quota ran out. Qwen and DeepSeek each got one question wrong. It was the same question, with the same wrong answer. Luna hadn't been given that question in its exam, so I asked it separately, with no tools and no web search. It chose the same wrong answer too. I checked the marked answer against Anthropic's documentation and it holds. The question turns on a detail of Anthropic's own API that changed earlier this year, and all three models know half of it. That's a useful result in two ways. A question that outside models disagree with is exactly the one worth re-checking, and this one survived the check. And the trap in it works: knowing roughly how something behaves isn't the same as knowing where the line is. The second Qwen run went wrong on every "choose two" question. It pressed a letter, looked at the page, decided nothing was selected, pressed again, and so took its own answer back. It was right that the page told it nothing. A chosen answer was highlighted visually, but nothing in the page's structure said which option was chosen. An agent reads that structure. So does a screen reader. So an AI agent failing my exam showed me that the site didn't tell blind users which answer they had picked. That's fixed now: a chosen answer is announced, and a build check fails if the highlight and the announcement ever come apart again. One more thing from the same run. When the agent got stuck it tried to force half an answer through with scripts. The server refused every attempt, because it checks each answer itself and doesn't trust what the page sends. I didn't plan that test, but I'm glad it passed. My first reaction was to do nothing. If someone wants to wear borrowed feathers on a practice leaderboard, that says more about them than about the board. But there are people on that board who earned their place the real way, and an agent run pushes every one of them down a row. So there is now a rule: I'm not going to describe how a run is recognised. Publishing that would only turn it into a puzzle. I will say how well it works, because I tested that too: it caught some of the runs above and not others. It can't tell whether a person had help beside them either, and I don't claim it can. It's a practice site, not a proctored exam, and I'd rather miss an agent than take a real player's result away. I also looked hard at the two callnames that started all this. Nothing shows that either of them is an agent, so both keep their place. In part 3 I promised that the Professional questions would follow the exam guide as strictly as Foundations does. That's done for the current guide, v1.0: The new Professional guide was due on 17 November , and the exam changes on 7 December . I'll rebuild to the new guide as soon as I have read it, well before the exam changes. Lost connections no longer lose answers. People study on trains. When a request fails because the connection dropped, the page now retries it before showing an error. The site counts its own faults. Server errors and page errors are counted per day by a fixed code, with nothing about who or where. Within a day that count pointed at seven false alarms, all caused by one else attached to the wrong if . A build check now fails if that pattern comes back. That brings the checks that run before every deploy to 82. Still free and unofficial: 950+ questions, 120 free practice questions a day without an account, and an 8-question demo if you just want a look. One question for you: if you run a leaderboard, a quiz or a coding challenge, have you tested what an agent scores on it? I'd like to hear what you found.