Tokenizer Comparison Sheet
Put the token counts your own models reported for one piece of text side by side, with characters per token and the monthly cost of the gap between them.
usage object on a response from that model, or by running that family’s tokenizer locally, and paste them in. The one row this page fills in by itself is marked as an estimate.Enter at least two counts to see the gap.
- This page's estimate (not a tokenizer)
- ~81 tokens · 4.37 chars/token · $24.30/mo
- Characters in the text
- 354
- Words in the text
- 63
Why the same text has several different token counts
A tokenizer is a vocabulary plus a set of merge rules learned from a training corpus. Two model families trained on different corpora end up with different vocabularies, so the same sentence genuinely is a different number of tokens depending on who you send it to. On English prose the families tend to land within about 10% of each other. On code, on structured data, and above all on non-Latin scripts, the gap gets much wider — a factor of two between families on the same paragraph is ordinary.
This matters in two places. The first is cost comparison: a headline per-million-token price is only comparable across models if the models agree on what a token is, and they do not. A model 20% cheaper per token that needs 30% more tokens for your text is more expensive. The second is context limits, where a prompt that fits comfortably in one family’s window overflows another’s even though the window is nominally the same size.
The fix is to compare on your own text rather than on a benchmark paragraph, which is what this sheet is for. Take a representative sample — a real prompt, with its real system message and its real retrieved context — send it once to each candidate, and copy the token counts out of the responses. That takes a few minutes and a few cents, and it gives you the only comparison that describes your workload rather than someone else’s.