Why does JSON mode cost more?

Every key, quote, brace, comma and bracket is itself tokens the model must generate. A field like "customer_name": "Jane" spends tokens on the key and punctuation on top of the value, so the same information costs more in JSON than in prose.

Does function calling have the same overhead?

Function/tool calling produces JSON arguments, so the output side carries similar structural overhead. Tool definitions in the prompt also add input tokens that this calculator does not include.

Is the schema sent anywhere?

No. Everything runs in your browser. The schema you paste is never uploaded, stored or logged.

Structured Output Token Overhead Calculator

Q: How is the overhead estimated?

The tool counts the structural (non-value) characters in your schema and converts them to tokens at roughly 4 characters per token, then compares total JSON tokens against your freeform baseline.

Structured output token overhead calculator

Forcing a model to return strict JSON, a fixed schema, or tool-call arguments is convenient — but every brace, quote, key and comma is output tokens the model has to generate and you have to pay for. This tool estimates how much extra you spend by demanding structured output versus letting the model answer in plain prose.

How it works

Output billing is per token, and structural characters are not free. Consider a single field:

"customer_name": "Jane Doe"

The value is Jane Doe, but you also pay for the key customer_name, two pairs of quotes, a colon and the surrounding braces and commas. Across a deeply nested schema with many keys, this “syntax tax” can add 20–60% to the output token count. The calculator counts the non-value (structural) characters in the schema you paste, converts them to tokens at roughly 4 characters per token, and compares the full JSON token count against the freeform baseline you provide.

Tips to cut structured-output cost

Use short, flat keys (id not customer_identifier) — keys repeat on every record and add up fast in arrays.
Avoid pretty-printing in production; request minified JSON with no extra whitespace or newlines.
For long lists, consider returning CSV or newline-delimited values when a strict schema is not required — far less punctuation per row.
Cache the prompt/schema where the provider supports prompt caching so the input side of the schema is cheaper, even though the output side still costs.