Counting tokens for DeepSeek models
DeepSeek publishes its tokenizer, so the counts on this page are exact rather than estimates. Both V4 models load by default and share one vocabulary, so the token count is the same and only the price changes. With a million-token context window on each, the context bar is the useful number when you are feeding in long documents or whole codebases.
Pricing and context windows
DeepSeek's input prices per million tokens, with the context window of each model.
| Model | Tokenizer | Price / 1M input | Context window |
|---|
| DeepSeek V4 Pro | deepseek-v4 | $0.44 | 1M |
| DeepSeek V4 Flash | deepseek-v4 | $0.08 | 1M |
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 DeepSeek tokenization works
DeepSeek V4 uses one byte-level BPE vocabulary across the whole family. The vocabulary ships in the model repositories as a standard tokenizer file, and this tool runs that exact file, so no approximation is involved. Byte-level BPE means any input encodes (there is no unknown token), and unusual characters simply cost more tokens.
DeepSeek's vocabulary carries dedicated entries for common Chinese sequences, so Chinese text tokenizes more efficiently here than with vocabularies built mainly around English.
Frequently asked questions
Is this DeepSeek's official tokenizer?
Yes. The tool loads the vocabulary DeepSeek publishes with the model. Nothing is estimated on this page: the file is fetched once and the text is tokenized in your browser, so the count matches what DeepSeek's API measures for the same text.
Do V4 Pro and V4 Flash count the same?
Yes. The same text is the same number of tokens on both. They share the vocabulary and the context window, so the compare table shows one count and two prices. That makes it easy to see what moving a workload from Pro to Flash would save.
Does DeepSeek tokenize Chinese differently?
Yes, and usually more efficiently than English-first vocabularies. Common Chinese words and phrases have their own entries, so a passage of Chinese costs fewer tokens here than with an OpenAI encoding. Paste a sample into the editor and switch models in the compare table to see the difference.
Does the count include the system prompt?
No, it counts only the text in the box. A real request also bills the system prompt, tool definitions, and the conversation history as input tokens. Count those separately and add them when you plan against a context window.
