Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: A token is the chunk of text a model reads and writes, roughly three quarters of an English word on average. Tokens are the unit of three things a PM manages: cost (providers bill per token, with output usually priced higher than input), speed (models generate output token by token), and memory (the context window is a token limit). Every prompt, document and answer length is therefore a product decision.
How to approach it
Define a token simply, give a rough conversion, then connect it to the three product consequences: cost, latency and context limits. Add one non-obvious point, such as different languages using different numbers of tokens for the same meaning.
A strong answer
What a token is. Models do not read words or letters. They read tokens, which are common chunks of text: a short word may be one token, a long or rare word several. As a rough rule in English, 100 tokens is about 75 words. A one-page document might be 500 to 700 tokens.
Why it matters, three ways.
Cost. Providers charge per token, separately for input (what you send) and output (what the model writes). As of 6 November 2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens: output is five times more expensive. A feature that sends a 4,000-token prompt and gets a 400-token answer costs about $0.012 per call. Multiply by calls per day and the shape of the bill becomes clear. Long system prompts, long chat histories and many retrieved documents all add input tokens on every single call. Price can also depend on prompt length: on the same date, Claude Haiku 5.5 input costs $0.10 per million tokens for prompts up to 100,000 tokens and $0.50 above that, so a design that crosses the threshold pays five times more for every input token.
Speed. Models produce output one token at a time. A 600-token answer at 80 tokens per second takes over seven seconds to finish. Longer answers are slower answers. OpenAI's latency guide puts it simply: cutting output tokens by half can cut latency by roughly half. Asking for concise output is a cost decision and a speed decision at once.
Memory. Each model has a context window, a maximum number of tokens it can consider at once (instructions, conversation, documents and the answer together). Long conversations and large documents eventually exceed it, and the product must decide what to summarise, drop or retrieve.
| Product choice | Token effect | Consequence |
|---|---|---|
| Longer system prompt | More input tokens every call | Higher cost, slightly slower first response |
| Keep full chat history | Input grows each turn | Cost and latency climb over a conversation |
| Retrieve 15 documents instead of 5 | Many more input tokens | Higher cost, possible distraction |
| Ask for detailed answers | More output tokens | Slower and more expensive |
Read diagram description
A short request split into nine illustrative tokens, then the three things tokens control: cost, where 4,000 input and 400 output tokens cost about 0.012 US dollars a call at October 2026 prices; speed, where 600 output tokens at 80 a second take 7.5 seconds; and memory, where one context window holds the system prompt, history, documents and answer together.
A non-obvious point. Token counts differ by language and by model. Text in Hindi, Tamil or other non-Latin scripts often needs noticeably more tokens than the same meaning in English, which makes the same feature more expensive and slower for those users. Petrov et al. (NeurIPS 2023) found that the same text translated into different languages can differ in token length by up to 15 times, even with tokenizers built for multilingual use. For a product priced the same in every market, that gap is a margin difference by country. Tokenizers also change between model generations; Anthropic documents that the tokenizer in Claude 4.7 and later models produces about 30% more tokens for the same text, as of 6 November 2026. Measure cost estimates on real traffic in each language; word counts mislead.
Price source: Anthropic rate card, checked 6 November 2026.
What interviewers probe next
"How would you estimate the monthly cost of a feature?" Tokens in and out per call, times calls per task, times tasks per month, times price, then adjust for retries and failures.
"How would you cut token usage?" Shorter stable prompts with caching, summarised history, fewer and better retrieved chunks, and shorter answers.
How it's scored
Weak: Defines a token vaguely or as a word, with no product consequence.
Hire: Defines tokens correctly and connects them to cost and context limits.
Strong Hire: Covers cost (with input versus output pricing), latency and context, gives a worked number, and notes language and tokenizer differences.
Common mistakes
Treating tokens as words.
Forgetting that the whole prompt is billed on every call, including the system prompt and the history.
Assuming token counts are the same across languages and models.