The AI Receipt

What's a token?

AI doesn't read in words - it reads in tokens, little chunks of text. One token is roughly ¾ of a word (about 4 characters). Short words like cat are one token; longer ones like unbelievable are three or four. A page of text is around 700 tokens.

Why it costs money

You pay per token, and there are two prices: input - everything you send in (your question, any files, and the AI's hidden setup) - and output, what the AI writes back. Output usually costs about 5× more than input.

Why files are the surprise

A 10-page PDF is several thousand tokens - and the AI re-reads all of them every time you ask a question about it. That's why "chat with my PDF" quietly gets expensive: you re-pay for the whole file on every message.

The trick that helps: caching

Prompt caching lets the AI "remember" a big file so it doesn't re-read it from scratch each time - repeat reads cost about 90% less.

Why there are three models

Haiku is cheapest and fast. Sonnet is the everyday workhorse. Opus is the smartest and priciest - worth it for hard reasoning, overkill for a quick summary.

The one thing to remember

For most tasks, the cost is what you feed in, not what the AI says back - and it's usually a lot cheaper than people fear. This tool just makes that visible.

These are ballpark figures, not a quote - your real cost depends on the exact text. They're built from Anthropic's own published anchors:

1 token ≈ ¾ of a word. A page of text ≈ 700 tokens. A ~12-page PDF ≈ 8,400 tokens. The AI's hidden setup instructions add roughly 1,800 tokens per request. A coding turn re-reads about 6,000 tokens of earlier context, and that grows as the session does.

Prices are the official Anthropic per-token rates, last verified on the date shown above. We round for readability and lean slightly high - so your real bill usually lands under the estimate, not over.

Source: Anthropic's pricing documentation (platform.claude.com).