Table Format Token Comparison
Renders your own table as CSV, TSV, markdown and three JSON shapes, and ranks them by estimated tokens at whatever row count you actually have.
Estimated tokens. The dearest, JSON, array of objects, indented, costs 227 — 2.87× as much for identical data.
| Format | Est. tokens | Per row | vs cheapest | |
|---|---|---|---|---|
| CSV | 79 | 13.8 | 1.00× | |
| TSV | 79 | 13.8 | 1.00× | |
| Markdown table | 101 | 15.8 | 1.28× | |
| JSON, header + rows | 111 | 16.8 | 1.41× | |
| JSON, array of objects | 137 | 26.8 | 1.73× | |
| JSON, array of objects, indented | 227 | 45.2 | 2.87× |
- Columns
- 4
- Rows in the sample
- 5
- Rows priced
- 5
- Cheapest, per month
- $2.37
- Dearest, per month
- $6.81
- Difference per month
- $4.44
- Difference per year
- $53.28
Estimated, not tokenized. Token figures here come from an approximation running in your browser: no tokenizer vocabulary is downloaded, because a real one is megabytes and nothing on this page fetches anything. It imitates how a byte-level BPE splits text — words, digit groups, punctuation runs, whitespace runs — but it has no merge table, so treat it as a planning number. The authoritative count is the usage object on a real API response.
The ranking surprises people, and it falls straight out of the arithmetic. JSON as an array of objects repeats every key on every row. With C columns and R rows you are paying for C × R copies of the column names, plus {, }, quotes, colons and commas around each one. CSV writes each column name exactly once and spends one comma per field. So the gap between them grows with row count and with how verbose your column names are — rename refund_rate to refundRatePercentage and only the JSON columns move.
Markdown tables sit in the middle and cost more than people expect, because | and the padding spaces around it are separate pieces to a tokenizer: a five-column row carries six pipes and the spaces between them, on every row, forever. CSV and TSV land close together — both spend exactly one separator per field, and which of the two wins depends on how your values merge with a comma versus a tab. The JSON "columns plus rows" shape is the interesting middle: it keeps the brackets and the quoting but names each column once, so it costs far less per row than an array of objects and the gap between those two widens with every row you add.
None of that settles which format to use. Structured formats buy you unambiguous parsing when values contain commas or newlines, and an array of objects is genuinely easier for a model to reference field by field in its answer. What the numbers above settle is the price of that convenience, so it becomes a decision rather than a default. If you are pushing thousands of rows into a prompt regularly, the honest move is usually CSV in the prompt and JSON in your code, with the column meanings explained once above the table instead of repeated on every line.