What this tool tells you
You rewrote a prompt and it reads better. The question nobody usually answers is whether it is cheaper, and by how much across a year. Paste both versions and the difference comes back in tokens, in dollars per month, and in dollars per year at your request volume.
Shorter is not automatically cheaper
This is the counterintuitive part, and the reason measuring beats estimating. Token count does not track word count reliably, because a tokenizer is a learned vocabulary rather than a formula.
Replacing a common phrase with a more precise but unusual one can add tokens while removing words — the common phrase had dedicated merges, the unusual one fragments. Restructuring a list into prose can go either way. Removing spaces to "save space" reliably makes things worse, because in most BPE vocabularies the leading space belongs to the word token.
The chars-per-token row in the comparison is the one to watch. If your rewrite is shorter in characters but the ratio dropped, you moved the text into denser territory and the saving is smaller than it looks. Watch it happen character by character in the tokenizer playground.
Deciding whether a change is worth shipping
Multiply the delta by your monthly request count — the tool does it for you — and apply a simple threshold:
- Under a few hundred dollars a year: ship it if the rewrite is better, ignore the saving. It is not worth the review cycle on its own.
- Low thousands: worth a deliberate change and a regression check.
- Above that: worth restructuring the prompt properly rather than trimming it — see the prompt builder.
System prompts deserve a lower threshold than user-facing prompts, because they are resent on every request by every user, and in a chat product on every turn as well. The system prompt analyzer prices them per paragraph and per year for exactly that reason.
This measures cost, not quality
Worth stating plainly, because the failure mode is predictable. A prompt that is 40% smaller because you removed two examples and a constraint may well perform worse, and the cost of a wrong answer usually exceeds the tokens you saved.
The order that works: rewrite for quality, verify quality with your evaluations, then use this tool to find out what the change cost or saved. Optimising for the number on this page without checking output is how teams ship a cheaper prompt that loses them money elsewhere.
A workflow that pays
- Run the original through the prompt optimizer first. Mechanical savings — indentation, minification — are free and carry no quality risk, so take them before touching wording.
- Rewrite the wording by hand for clarity.
- Diff the two here to see what the rewrite actually cost or saved.
- Evaluate output quality on both versions before shipping either.
Keep the losing version. Prompt changes get reverted more often than people expect, and having the token and cost figures already measured for both makes that conversation short.