Does Asking for JSON Cost More? Output Format and Your Token Bill

Format Is Tokens, and Tokens Are Money

When you ask an AI model to "return the result as JSON" or "format this as a Markdown table," you are not just choosing how the output looks — you are choosing how many output tokens the model generates. And since output tokens are the most expensive part of most API bills (often 4–5x the input rate), the format you request has a direct, measurable effect on cost.

The difference is not huge on a single call, but it compounds. If every response in a high-volume pipeline carries 20% of avoidable structural overhead, you are paying a 20% tax on your entire output bill. For a coding agent generating structured data thousands of times a day, that adds up.

Comparing the Formats

Every format spends tokens on structural characters — the "syntax tax" that carries no semantic payload. Here is roughly how the common ones stack up for the same data:

• Plain text / prose: the leanest for freeform content — no delimiters, no keys. But it is not machine-parseable, so it only works when a human reads the output.
• Markdown: very light. A few characters for headers, bullets, and table pipes. Excellent balance of structure and low overhead.
• JSON: moderate overhead. Every value carries quotes, keys are repeated on every object, and braces and commas add up — especially in arrays of similar objects where the same keys repeat hundreds of times.
• XML: the heaviest. Every field is wrapped in an opening and closing tag, so tag names are paid for twice per value. Verbose tag names multiply the cost.

A rough rule of thumb: for the same structured data, XML often costs noticeably more output tokens than JSON, and JSON costs more than an equivalent Markdown table or CSV. The gap widens as the data gets more repetitive, because repeated keys and tags are pure overhead.

A Worked Example

Say you ask a model to return 100 records, each with five fields. In JSON, every one of those 100 objects repeats all five field names, plus quotes, colons, commas, and braces. That is 500 repetitions of your key names — pure structural cost. The same data as a CSV names each field once in a header row, then sends only the values. For repetitive tabular data, the CSV can use a fraction of the output tokens.

At a premium output rate — say $25 or $30 per million tokens — trimming thousands of output tokens per call across a high-volume workload turns into real monthly savings. The model did the same reasoning either way; you just paid less to transmit the answer.

Choosing the Cheapest Format That Works

Cost is not the only consideration — you need a format your downstream code can actually consume. The practical guidance:

• Tabular, repetitive data? Prefer CSV or a Markdown table over JSON — you avoid repeating keys on every row.
• Need machine parsing with nesting? Use JSON, not XML — same structure, fewer tokens.
• Output is for a human? Markdown or plain prose is both cheaper and more readable than forcing JSON.
• Use short, consistent keys. If you must use JSON, "id" beats "identifier" when it repeats thousands of times.

Don't force structure you don't need. A common anti-pattern is demanding elaborate JSON for output a human will read once — you pay the syntax tax for nothing. Match the format to the consumer.

Want to see how output-token savings translate to dollars at your volume? Plug your expected output size and call frequency into the AI Cost Estimator and watch how a leaner format changes the monthly total.