This quiz is interactive and needs JavaScript. Here are the questions for reference:
1. Which of the following counts toward a model's context window?
- A) Only the visible user and assistant text
- B) The system prompt, tool definitions, tool results, images, and generated output including extended thinking
- C) Everything except tool results, which are billed separately
- D) Only
input_tokens; cached buckets are excluded
2. ⚠ (point-in-time) A task needs a single request to hold a very large document set that a 200k-token model cannot fit. Which grouping of models offers the ~1M-token window?
- A) Haiku 4.5, Sonnet 4.5, Opus 4.5
- B) Opus 4.8, Sonnet 5, Fable 5, Sonnet 4.6
- C) Only Opus-family models
- D) Any model, once you send the 1M beta header
3. On a prompt-caching request, what are the cost multipliers (× base input) for a cache read, a 5-minute cache write, and a 1-hour cache write?
- A) 0.1× read, 1.25× (5m) write, 2× (1h) write
- B) 0.5× read, 1× (5m) write, 1.5× (1h) write
- C) 0.1× read, 2× (5m) write, 4× (1h) write
- D) 0.25× read, 1.25× (5m) write, 1.25× (1h) write
4. What are the breakpoint and lookback limits for prompt caching?
- A) Unlimited breakpoints, 4-block lookback
- B) Max 4 explicit
cache_control breakpoints, 20-block lookback per breakpoint - C) Max 20 breakpoints, 4-block lookback
- D) Exactly 1 breakpoint, whole-prompt lookback
5. A prompt is shorter than the model's minimum cacheable length. What happens?
- A) The API returns a 400 error
- B) It is processed normally without caching, and no error is returned
- C) It is cached anyway at a higher write price
- D) The request is silently dropped
6. You are running a large classification job through the Message Batches API and want to layer prompt caching on top. Which cache TTL should you choose, and why?
- A) The 5-minute TTL, because it is cheaper to write
- B) The 1-hour TTL, because the 5-minute cache would expire before the async batch job runs
- C) Either — TTL has no effect on batch jobs
- D) No TTL; caching is disabled inside the Batches API
7. In the context-diagnosis map, a system that misroutes to the wrong tool most likely has a problem with which of the following?
- A) The system prompt
- B) Lost in the middle
- C) Tool descriptions
- D) Progressive summarisation
8. A confidence-based extraction pipeline auto-processes fields above 90% confidence. Before reducing human review, what is the critical validation step?
- A) Trust the model's self-reported confidence once it averages above 90%
- B) Sample high-confidence items uniformly at random
- C) Analyse accuracy by document type and field to confirm consistent performance across all segments, not just in aggregate
- D) Raise the auto threshold to 99% and stop auditing
9. Prompt caching is ZDR-eligible, but the MCP connector is not.
10. On Claude 4.5 and newer models, when input plus max_tokens would overflow the window, the API rejects the request with a 400 error before generating anything.
11. Distinguish context rot from lost in the middle in one sentence each.
12. Name the two tags the API injects for context awareness on models like Sonnet 5 and Haiku 4.5, and state what newer Opus models (4.7+) use instead.
13. A 48-turn support session is approaching the context window. One resolved billing issue and one still-active shipping issue are in the history. Following the structured-summary playbook, how should the transcript be laid out to avoid progressive-summarisation detail loss?
14. An agent on Sonnet 5 gives inconsistent answers to a general policy question across sessions — but tool routing and deep-context recall are fine. Which single fix does the context-diagnosis map point to?