Claude Certified Architect – Foundations Practice Question 50: Token Counting Before Send
Claude Certified Architect, Foundations · Test 5 · question 10 of 10
A document-analysis service assembles requests from a system prompt, tool definitions, a chat history and user-uploaded PDFs. Some requests exceed the model's context window and fail, and finance wants to estimate the input cost of large requests before committing to them. The team wants a way to learn the input size of the exact request, without generating a response and without paying for the check. What should they use?
Show answer and explanation
Answer:
The token counting endpoint accepts the same structured input as message creation, including system prompts, client tools, images and base64 PDFs, and returns the total input tokens without creating a message. It is free to use, subject to its own requests-per-minute limits, which are separate from message creation limits. The count is an estimate that can differ slightly from actual usage, and it is tokenizer-specific, so you should count against the model you plan to use (newer models use a tokenizer that can produce roughly 30 percent more tokens for the same text). That makes it suitable for checking a request fits the context window and for estimating cost.
A is wrong because characters divided by four is a rough heuristic, not a count, and it ignores tool definitions and PDFs. B is wrong because it actually runs billable inference on every probe, and it also still starts generation. C is wrong because a generic tokenizer does not match Claude's tokenizer and cannot account for images, PDFs or tool definitions, and counts differ between Claude model generations.
Reference: Token counting