Token counter On-device
Tokens and cost per model — and why Arabic costs more.
Paste on the left to start.
Paste above to start.
Everything runs on your device. Nothing you paste is sent anywhere, and it's cleared when you close the tab.The exact GPT and Llama tokenizers are downloaded once (about 3.5 MB) and kept offline.
Paste text and see how many tokens it uses and what it costs on each model. GPT and Llama counts use the real tokenizers, downloaded once and kept on your device; Claude and Gemini have no published tokenizer, so those figures are labelled estimates. An Arabic and English split shows why the same text can cost more in Arabic. Nothing you paste is sent anywhere.
About this tool
What you provide
- Text pasted or opened from a .txt, .md, .json, .csv or .log file up to 2 MB.
What you get
- Token counts and cost per selected model, with an Arabic, English and other script split.
- A JSON summary and, for exact tokenizers, the individual tokens.
Limits
- Claude and Gemini counts are estimates; no tokenizer is published for them.
- Prices are a dated snapshot and may change.
- Exact GPT and Llama counting needs a one-time download of about 3 MB.
How to use it
-
Paste your text
Paste into the Input box, or open a text file.
-
Choose models
Pick up to four models; each shows Exact or Estimate.
-
Read the breakdown
See tokens, cost and how much of each model's context the text fills.
Frequently asked questions
Why does Arabic use more tokens than English?
Most tokenizers were built mostly on English text, so Arabic words are split into more, smaller pieces. The same sentence can cost noticeably more in Arabic, and diacritics add further tokens.
Are the counts exact?
For GPT and Llama models the tokenizer is the real one, so the counts are exact. Claude and Gemini have no published tokenizer, so their figures are estimates and are marked as such.
Is my text uploaded?
No. Counting happens on your device. The only download is the tokenizer data, fetched once and then kept for offline use.