Give a quick, honest estimate and the cheapest way to get the same result. These are estimates; say so. Exact counts come from the API's token counter.
Estimate
- English prose: about 4 characters per token, roughly 750 words per 1,000 tokens. A full page of text is 500–700 tokens.
- Code, JSON, logs: about 3 characters per token.
- Cyrillic, CJK and other non-Latin scripts: about 1.5–3 characters per token, so the same text costs noticeably more than in English.
- Images: on the order of 1,500 tokens each at typical sizes. PDFs: the text, plus an image's worth for pages read as images.
Give a range (for example "about 18–24K tokens"), and compare it with the context window: 1M tokens on current Opus, Sonnet and Haiku 5.5 models, 200K on older Haiku.
Explain the real cost
In a conversation, everything already in it is re-sent with every new message. A 20K-token document pasted early is paid for again on each of the next 30 messages, about 600K tokens in total. That, not the first send, is what uses up limits.
Ways to cut it, cheapest first
- Send only the part that matters: the section, the function, the failing lines of a log.
- Strip what carries no meaning: minified code, base64, repeated log lines, boilerplate.
- Summarize a long document once, then continue from the summary.
- When a conversation is long, start fresh from a handoff (the
fresh-startskill). - In Claude Code: read files with offset/limit, let the
context-optimizerscout subagent do wide exploration on Haiku, and run/cco auditto see where a session's tokens actually went.