tokenmeter

Count tokens and estimate cost for prompts before you send them, from the command line or as a CI budget gate.

Prompt templates grow, a few-shot example gets added, a retrieved context balloons, and suddenly every call costs more than you thought. tokenmeter gives you the exact token count and a dollar estimate up front, for a single prompt or a whole directory of templates.

$ tokenmeter count prompts/system.txt --model gpt-4o
input            in tok   out tok   cost (USD)
prompts/system.txt  812         0   $0.002030

$ tokenmeter count prompts/ --model gpt-4o-mini --json

Install

$ pip install tokenmeter-cli                 # from PyPI, once released
$ pip install git+https://github.com/jmweb-org/tokenmeter   # latest, available now

Token counting is exact for the supported OpenAI encodings via tiktoken.

Usage

$ tokenmeter count system.txt -m gpt-4o          # one file
$ tokenmeter count prompts/ -m gpt-4o-mini       # every text file in a directory
$ cat prompt.txt | tokenmeter count - -m gpt-4o  # standard input
$ tokenmeter count p.txt --output-tokens 500     # include an assumed completion
$ tokenmeter models                              # list models and prices

As a budget gate

Fail a build when a prompt set would cost more than you allow:

$ tokenmeter budget prompts/ --model gpt-4o --max-cost 0.05

- run: tokenmeter budget prompts/ --model gpt-4o --max-cost 0.05

Cost model

Counts are real tokens. Cost multiplies tokens by a per-model rate from a small, dated price table (tokenmeter models prints it with its "as of" date). By default only input tokens are counted; pass --output-tokens N to add an assumed completion length to the estimate. Prices change, so treat the dollar figures as estimates and update the table when they move.

Exit codes

License

MIT. See LICENSE.