Counting tokens for Mistral models
Mistral publishes its tokenizer, so the counts on this page are exact. Large, Medium 3.5, and Small load together and share one vocabulary, so the compare table shows a single token count against three prices: the fastest way to see what a prompt costs as you move up and down the lineup.
Pricing and context windows
Mistral input prices per million tokens, with the context window of each model.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| Mistral Large | mistral-3-4 | $0.50 | 262K |
| Mistral Medium 3.5 | mistral-3-4 | $1.50 | 262K |
| Mistral Small | mistral-3-4 | $0.15 | 262K |
Prices and context windows are curated metadata refreshed periodically from the providers and OpenRouter, so treat them as guidance rather than a quote. Token counts are always produced by the real tokenizer.
How Mistral tokenization works
One vocabulary covers the Mistral 3 and 4 generations. Mistral ships a byte-level BPE vocabulary shared across its current models, so Large, Medium, and Small always agree on a count. The tool loads that published file directly, which makes these numbers exact rather than approximations.
On plain English prose Mistral's vocabulary is slightly less compact than OpenAI's (in our fixed 500-word sample it produced the highest count of any family we measured), so budget a little more headroom when porting a prompt from GPT.
Frequently asked questions
Which Mistral models share a tokenizer?
All the current ones: Large, Medium 3.5, and Small use the same vocabulary. The same text is the same number of tokens on each, so a single Mistral counter covers the family. Price and context window are what differ between them.
Is this Mistral's official tokenizer?
Yes. The tool loads the vocabulary Mistral publishes with its models. It is fetched once and run in your browser, so the count is exact and your text is never uploaded.
How do Mistral counts compare with GPT?
Slightly higher on English prose, by roughly one to two percent. Our measured 500-word sample came to 581 tokens with Mistral's vocabulary against 572 with o200k_base. The gap widens on code and non-English text, so measure your own content rather than assuming parity.
Does the count include the system prompt?
No, it counts only the text in the box. System prompts, tool definitions, and earlier messages are billed as input tokens on a real request. Count them separately and add them when you plan against a context window.
