How to Count JSON Tokens for LLMs

Quick answer: LLMs bill and limit by tokens, and JSON is token-expensive because every brace, quote, comma, and repeated key counts. To estimate a payload's cost, paste it into the JSON Token Estimator, which approximates token counts and compares JSON against YAML so you can pick the cheapest representation for a prompt or context window.

A token is a chunk of text — roughly ¾ of a word in English, but for structured data it's often a single character or symbol. Models don't see your JSON as data; they see a sequence of tokens. That means the punctuation of JSON is not free:

{"user":{"name":"Dana","active":true}}

Every {, }, ", :, and , is tokens the model must process — and that you pay for on metered APIs. The same data carries a lot of structural overhead beyond the actual values.

• Quotes on every key and string — "name" costs more tokens than a bare name.
• Repeated keys — an array of 1,000 objects repeats every key 1,000 times.
• Braces, brackets, and commas at every level of nesting.
• Whitespace — pretty-printed JSON adds indentation tokens. (Minify before sending to a model.)

This is why feeding large JSON into a prompt burns through a context window fast and inflates cost.

1. Paste your JSON into the JSON Token Estimator.
2. Read the approximate token count for the payload as-is.
3. Compare the JSON estimate against the YAML estimate the tool shows — YAML often uses fewer tokens because it drops quotes and braces.
4. Decide whether the payload fits your model's context budget, or needs trimming.

Token counts are approximate — each model family uses its own tokenizer — but a good estimate is enough to catch a payload that's about to blow your budget.

• Minify it. Strip pretty-printing whitespace before sending — use the JSON Formatter in minify mode. The data is identical; the token count drops.
• Send only what's needed. Use Minimal Mode to include or exclude paths and produce a smaller subset, instead of pasting the whole document.
• Consider a more compact format. For some prompts, YAML or a tabular form conveys the same data in fewer tokens. The estimator's side-by-side comparison shows the difference.
• Drop redundant nesting. Flatten where the structure isn't load-bearing for the model's task.

For background on why compact vs readable JSON matters, see Pretty vs Minified JSON, and for cleaning up the JSON that models return, see Fixing Broken LLM JSON.

How many tokens is my JSON? It depends on the content and the model's tokenizer, but JSON's quotes, braces, and repeated keys add substantial overhead. Estimate it with the JSON Token Estimator.

Does JSON use more tokens than YAML? Usually yes. YAML omits most quotes and braces, so the same data often tokenizes smaller — which is why the estimator compares the two.

Does minifying JSON reduce tokens? Yes. Removing pretty-print whitespace lowers the token count without changing the data. Minify before putting JSON in a prompt.

Why does my JSON exceed the context window? Large arrays repeat every key per element and nesting adds punctuation tokens. Trim to the needed paths with Minimal Mode or send a more compact format.