Paste any text or code and instantly see the token estimate plus the price on Opus, Sonnet, and Haiku — as input and as output, including cache rates. No signup, nothing uploaded.
Paste something to count. Rough rule: 1 token ≈ 4 characters of English, less for code.
100% client-side · nothing uploaded · calibrated for prose and code
Per million tokens, July 2026. Output always costs ~5× input; cache reads cost 90% less than fresh input — the two facts that explain most Claude bills.
Input $15 / MTok · Output $75 / MTok
Cache write $18.75 · Cache read $1.5
The heavyweight — for the hardest reasoning and agentic work.
Input $3 / MTok · Output $15 / MTok
Cache write $3.75 · Cache read $0.3
The default — best cost/intelligence balance for coding.
Input $1 / MTok · Output $5 / MTok
Cache write $1.25 · Cache read $0.1
The economy tier — classification, extraction, high-volume calls.
English prose averages 3.5–4 characters (≈0.75 words) per token. 'The quick brown fox' is 4 tokens; this whole sentence is about 12.
Brackets, operators, and camelCase identifiers split into more tokens per character — budget 25–40% more than prose. This counter detects code and adjusts.
Every call bills the entire prompt (input) plus the response (output) — and in a conversation, the whole history is re-sent as input each turn. Long chats compound.
Repeated context (system prompts, codebases) can be cached: 25% extra to write, 90% off to read. It's why Claude Code can hold your repo in context affordably.
count_tokens endpoint — it runs the real tokenizer server-side. For planning, budgeting, and prompt-size sanity checks, this estimator's ±5% is plenty.The ClaudeThings kits are token-efficient by design — tight skills, scoped agents, and commands that get more done per call, for engineering and marketing.
Battle-tested Claude skills & workflows for builders.
Turn Claude into your growth team.
Or grab all 89 agents, 103 skills & 181 commands in the Complete Bundle →