token-diet

See what your JSON costs.
Then put it on a diet.

Most apps ship LLMs pretty-printed JSON — the most expensive way to send structured data. Paste yours below: get the real token count and dollar cost, watch token boundaries light up, then cut it losslessly with JTF.

tokenizer
sample
loading tokenizer…
input.json valid
tokens json-pretty
0
est. cost / Sonnet in
$0.00
token chips each chip = 1 token of the current encoding
encodings, by token count lower is cheaper · baseline = json-pretty
cost across models current encoding · input tokens

Honest measurement. Counts come from a real GPT-family tokenizer (gpt-tokenizer, cl100k_base / o200k_base) running in your browser — never estimates. Claude tokenizes differently; treat the count as a close GPT-family proxy, not exact Anthropic counts. For exact Anthropic counts use token-diet bench --api. Prices are live from LiteLLM (USD / 1M tokens, priced as input tokens), cached locally and refreshed daily — prices: snapshot. JTF is the project's own lossless format — these are the true encoder's numbers.