AI Foundations: AI Engineering Basics Practice Questions

Only the 3,000-token system prompt is cached. It saves ($5.00 − $2.50) per 1M tokens: 3,000 × $2.50 / 1,000,000 = $0.0075 per request. Over 10,000 requests that is $75 per day. $37.50 halves the differential, $125 applies the saving to all 3,500 tokens at $5.00, and $175 is the full daily input cost.

The correct answer is C ($270.00). Input tokens per month: 800 × 50,000 = 40,000,000 tokens = 40M tokens. Input cost: 40 × $3.00 = $120.00. Output tokens per month: 200 × 50,000 = 10,000,000 tokens = 10M tokens. Output cost: 10 × $15.00 = $150.00. Total monthly cost: $120.00 + $150.00 = $270.00. A ($27.00) is off by a factor of 10 — a common error when miscounting zeros in the per-million conversion. B ($120.00) only accounts for input cost and ignores output cost entirely. D ($285.00) results from incorrectly pricing output tokens (e.g., applying $13.50/M instead of $15.00/M).

HTTP 429 is the standard 'Too Many Requests' status. It means the application has exceeded its rate limits (Requests Per Minute and/or Tokens Per Minute). The industry-standard fix is exponential backoff with jitter: wait an increasing delay before each retry, which spreads load and avoids thundering-herd problems. The invalid-API-key answer is wrong because an invalid key returns HTTP 401 (Unauthorized), not 429. The malformed-payload answer is wrong because a malformed request returns HTTP 400 (Bad Request), not 429. The deprecated-model answer is wrong because a deprecated model typically returns a 404 or a specific error message in the response body, not a 429.

Prompt caching is built for this scenario. When a large, static prefix (like a system prompt or long document) is reused across many requests, both Anthropic's and OpenAI's caching mechanisms let the model skip reprocessing the cached tokens. That cuts input token costs and improves latency. The Batch API answer is wrong because the Batch API is for asynchronous, non-real-time jobs (e.g., overnight classification). It does not help a live, interactive document assistant that needs low-latency responses. The function-calling answer is wrong because tool definitions still count as input tokens and are not a caching mechanism, so they don't remove the cost of the system prompt. The model-versioning answer is wrong because pinning an older model may lower per-token cost slightly, but it does not stop the repeated prompt from being reprocessed, and it may sacrifice quality.

The Batch API is built for large-scale, asynchronous workloads. The 50% cost discount is correct: OpenAI's Batch API costs half as much as synchronous pricing, which makes it very cost-effective for bulk jobs. Asynchronous processing is correct: jobs are submitted as a file, processed asynchronously (up to 24 hours), and results are retrieved later, which suits a weekend batch run. The claim of synchronous processing with results in about one second is false, because batch latency can be up to 24 hours. The claim that prompts are split automatically to fit context windows is false, because the Batch API does not manage context windows. The claim that results are streamed over server-sent events is false, because results are written to an output file for retrieval. The statement about offline, non-time-sensitive workloads is conceptually true, but it is treated here as not being a distinct technical 'characteristic'. The cost discount and asynchronous processing are the precise, documented features.