Claude Certified Associate – Foundations Exam

Who can sit it, and how registration works

Exam-day rules to know before you book

• Treat 48 hours as the change deadline. The exam guide (version 1.0) lets you cancel or reschedule up to 24 hours before the appointment; Anthropic’s certification FAQ says nothing can be cancelled or rescheduled inside 48 hours. Both still said so when we read them on Oct. 1, 2026, and only the 48-hour deadline is safe under both. A change inside the window, a no-show, or arriving after the permitted late-arrival window forfeits the fee. After rescheduling, check that the confirmation email arrived. Cancelling in Pearson does not refund you; to ask for a refund, email Anthropic’s certifications support.
• Your name. The name on your registration and on your government-issued photo ID must match exactly, and a mismatch you did not report in advance means Pearson will not let you test. Corrections go to certifications support at least 24 hours ahead and typically take 24–48 business hours, so in practice ask several business days before.
• Nothing but you. The exam is closed book. Anthropic’s exam policy bans using AI products or services during the exam, Claude included; browser translation is not allowed, and the exam exists in English only. Online, you stay in webcam view for the whole session with a desk clear of notes, books, phones and second monitors.
• Where you can test online. OnVUE refuses IDs from Belarus, Cuba, North Korea, Russia, Syria and restricted regions of Ukraine; those candidates book a test center. Pearson suspended delivery to residents of Iran, online and at test centers, from Sept. 8, 2026.
• The confidentiality agreement. You accept it before the first item; declining ends the session and the fee is not refunded.

The four Claude certifications

All four are partner-only, allow 120 minutes and pass at 720 on the 100–1,000 scale; the fees are before any tier discount. Only Developer and Architect passes count toward a firm’s Claude Partner Network eligibility; the Associate is the one written for people who do not code. Our comparison of the four exam guides covers their blueprints, the Architect exam’s scenario draw and retake costs.

What it costs at your partner tier

The list fee is $99 an attempt, and every attempt, retakes included, is charged at your organization’s Claude Partner Network tier. The discount is applied automatically at checkout.

The dollar amounts are our arithmetic on Anthropic’s list fee and discounts. One consequence for Global Premier staff: the price is set when you pay at registration, and a registration stays valid for five years, so registering by Dec. 31, 2026 should lock in that one free attempt even if you sit it in 2027. That is our reading of two FAQ answers rather than a rule Anthropic states, and the FAQ does not say whether the Dec. 31 date applies to the purchase or the exam; if your timing depends on it, confirm with partner support. A forfeit (a late change, a no-show, arriving after the late-arrival window) loses whatever you paid.

If you fail: the earliest retake dates

The wait grows with each failure, you can sit the exam at most four times in any rolling 12 months, and when you register again in Partner Academy, Pearson only offers dates after the wait. If every attempt fails and you book the first date Pearson offers, counting from a first attempt on Oct. 15, 2026:

So all four attempts can fall within 134 days; the guide sets no wait after a fourth failure, and the rolling limit is what holds a fifth back. Limits apply per exam, so failing the Associate does not stop you registering for another Claude exam. Your score report shows the percentage you answered correctly in each domain; those percentages do not decide the result, and Anthropic offers them for choosing what to review before a retake.

What the 60 items cover

Anthropic weights seven domains from a job task analysis, each percentage being the approximate share of scored items. Over 60 items that is roughly 13 on judging Claude’s output and 6 on troubleshooting. Anthropic’s free prep course on Partner Academy has seven skill lessons whose titles track the seven domains; the table pairs them, with one objective per domain from the guide that shows the level it expects.

Item counts are our arithmetic (weight × 60, rounded so the column sums to 60); Anthropic publishes only the percentages. The lesson pairing is ours too, by title; the one lesson not named after its domain, Claude Platform & Model Foundations, covers Domain 3’s features, models and context. The course runs 389 minutes including an 8-minute summary, and by our count its time tracks the weights, with one tilt: it gives Product and Model Selection the most minutes per exam item (about 8) and Troubleshooting the fewest (about 5). The guide’s own preparation list points to Anthropic’s documentation for features such as Projects, Artifacts, Memory, Skills and Code Execution.

What the guide’s three sample questions reward

The guide prints three sample items, from Domains 2, 3 and 6, and says they are not drawn from the live item bank. Each keyed answer gets the task done while handling its one constraint: verify a cited subsection against the regulation before the summary reaches the compliance team; match a faster, lower-cost model to high-volume reply drafts; remove or anonymize customer identifiers before uploading the spreadsheet.

The nine wrong options fall into three kinds, each matching the wording of the guide’s own rationale. Taking it at face value: send the summary because Claude sounded confident, or ask Claude to rate its own confidence (the guide: self-reported confidence is not a reliable accuracy signal); upload the spreadsheet because the analysis is internal. A move that does not touch the problem: reword the summary to sound formal (reformatting does not address correctness); disable all product features or switch AI platforms (neither addresses the trade-off); tell Claude not to retain the file (that does not satisfy the policy control). Overcorrecting: use the most capable, most expensive model for every reply (it wastes the cost and latency budget), or skip the analysis (abandoning the task is unnecessary when anonymization enables it). The grouping is our reading of three items, not a rule Anthropic states.

Our practice questions are written to the same seven domains; the chart below shows how they split.