· AI Engineering · 3 min read
Exam At a Glance
| Item | Detail |
|---|---|
| Format | Multiple-choice and scenario-based questions |
| Question count | ~60 questions |
| Time limit | 120 minutes |
| Passing score | 75% (~45 of 60 correct) |
| Delivery | Online, proctored |
| Credential validity | 2 years |
Question Style — Scenario-Based, Not Recall
The most important thing to understand about this exam: it tests judgment, not memorization.
The majority of questions give you a system design scenario and ask you to select the best architectural choice. Pure recall questions (“what is the minimum cache block size?”) exist but are a small minority.
Example scenario question:
A legal tech company processes the same 100-page contract template across 500 client queries per day. The system prompt is 5K tokens and the template is 40K tokens. What is the most cost-efficient architecture?
A) Use RAG — chunk the contract and retrieve relevant passages per query
B) Pass the full contract in context with prompt caching enabled
C) Use Claude Opus to maximize analysis accuracy
D) Summarize the contract once and pass the summary in context
(Answer: B — the contract fits within the 200K context window, and prompt caching eliminates re-processing costs on all 499 subsequent queries. RAG adds unnecessary complexity and latency.)
What makes a question hard: The wrong answers are usually plausible. Option A above looks reasonable — RAG is a legitimate pattern. The exam tests whether you know when RAG is worse than in-context.
Domain Weights
Study time should roughly match exam weight:
| Domain | Weight | Recommended Study Share |
|---|---|---|
| Domain 1: Model Selection and Capabilities | 15% | 1 week |
| Domain 2: Prompt Engineering | 25% | 2 weeks |
| Domain 3: Context, Memory, and Caching | 20% | 1.5 weeks |
| Domain 4: Tool Use and Multi-Agent Systems | 25% | 2 weeks |
| Domain 5: Safety, Compliance, and Deployment | 15% | 1 week |
Domains 2 and 4 together account for 50% of the exam. If you have limited study time, prioritize prompt engineering and tool use / agents.
Time Management
120 minutes for ~60 questions = 2 minutes per question average.
Scenario questions often require reading a paragraph of context before evaluating four options. Budget 3 minutes for complex scenarios and 1 minute for straightforward recall questions. Flag uncertain questions and return to them.
Do not spend 10 minutes on a single question — skip and return.
What You Are Allowed to Use
The exam is closed-book and proctored. You may not reference the Anthropic documentation, code editors, or external resources during the exam. This is why hands-on practice matters more than document reading — you need the knowledge to be internalized, not looked up.
Common Failure Patterns
Based on exam feedback, candidates most commonly lose points on:
- Prompt caching mechanics — confusing minimum token thresholds, cache TTL, or placement rules
- Multi-agent vs. single-context decisions — over-engineering with agents when a single context window suffices
- Tool description quality — underestimating how much the description drives tool selection behavior
- Output guardrails — knowing what to validate, not just that you should validate
- Extended thinking scope — using it on tasks that don’t benefit (adds cost/latency, no accuracy gain)
Scheduling and Renewal
Check the Anthropic certification page for current scheduling details. The credential expires after 2 years. Renewal requires passing a refreshed exam, as model APIs and best practices evolve significantly over that period.
Next Steps
Proceed to How to Build Your Study Plan to get an 8-week week-by-week preparation roadmap.