`tiktoken-rs`

Rust library for tokenizing text with OpenAI models using tiktoken.

This library provides a set of ready-made tokenizer libraries for working with GPT, tiktoken and related OpenAI models. Use cases cover tokenizing and counting tokens in text inputs.

This library is built on top of the tiktoken library and includes some additional features and enhancements for ease of use with Rust code.

Supported OpenAI models include GPT-5.4, GPT-5, GPT-4.1, GPT-4o, o1, o3, o4-mini, and gpt-oss models.

Scope: This crate is focused on OpenAI tokenizers (tiktoken). For non-OpenAI models (Llama, Gemini, Mistral, etc.), use the HuggingFace tokenizers crate.

Examples

For full working examples for all supported features, see the examples directory in the repository.

Usage

Add the library to your project:

cargo add tiktoken-rs

Use the examples below in your Rust code.

Counting tokens

use tiktoken_rs::o200k_base;

let bpe = o200k_base().unwrap();
let tokens = bpe.encode_with_special_tokens(
  "This is a sentence   with spaces"
);
println!("Token count: {}", tokens.len());

For repeated calls, use the singleton to avoid re-initializing the tokenizer:

use tiktoken_rs::o200k_base_singleton;

let bpe = o200k_base_singleton();
let tokens = bpe.encode_with_special_tokens(
  "This is a sentence   with spaces"
);
println!("Token count: {}", tokens.len());

Encoding text

CoreBPE::encode mirrors upstream tiktoken and returns a Result. Propagate or unwrap the result before using the tokens:

use tiktoken_rs::o200k_base;

let bpe = o200k_base().unwrap();
let allowed = bpe.special_tokens();
let (tokens, last_piece_token_len) = bpe.encode("hello <|endoftext|>", &allowed).unwrap();

The generic encode_as and count helpers also return Result. The crate requires Rust 1.85 or newer.

Counting max_tokens parameter for a chat completion request

use tiktoken_rs::{get_chat_completion_max_tokens, ChatCompletionRequestMessage};

let messages = vec![
    ChatCompletionRequestMessage {
        content: Some("You are a helpful assistant that only speaks French.".to_string()),
        role: "system".to_string(),
        ..Default::default()
    },
    ChatCompletionRequestMessage {
        content: Some("Hello, how are you?".to_string()),
        role: "user".to_string(),
        ..Default::default()
    },
    ChatCompletionRequestMessage {
        content: Some("Parlez-vous francais?".to_string()),
        role: "system".to_string(),
        ..Default::default()
    },
];
let max_tokens = get_chat_completion_max_tokens("o1-mini", &messages).unwrap();
println!("max_tokens: {}", max_tokens);

Counting max_tokens parameter for a chat completion request with async-openai

Enable the async-openai feature in your Cargo.toml file.

# #[cfg(feature = "async-openai")]
# {
use tiktoken_rs::async_openai::get_chat_completion_max_tokens;
use async_openai::types::chat::{
    ChatCompletionRequestMessage, ChatCompletionRequestSystemMessage,
    ChatCompletionRequestSystemMessageContent, ChatCompletionRequestUserMessage,
    ChatCompletionRequestUserMessageContent,
};

let messages = vec![
    ChatCompletionRequestMessage::System(ChatCompletionRequestSystemMessage {
        content: ChatCompletionRequestSystemMessageContent::Text(
            "You are a helpful assistant that only speaks French.".to_string(),
        ),
        name: None,
    }),
    ChatCompletionRequestMessage::User(ChatCompletionRequestUserMessage {
        content: ChatCompletionRequestUserMessageContent::Text(
            "Hello, how are you?".to_string(),
        ),
        name: None,
    }),
];
let max_tokens = get_chat_completion_max_tokens("o1-mini", &messages).unwrap();
println!("max_tokens: {}", max_tokens);
# }

tiktoken supports these encodings used by OpenAI models:

Encoding name	OpenAI models
`o200k_harmony`	`gpt-oss-20b`, `gpt-oss-120b`
`o200k_base`	GPT-5 series, `o1`/`o3`/`o4` series, `gpt-4o`, `gpt-4.5`, `gpt-4.1`, `codex-*`
`cl100k_base`	`gpt-4`, `gpt-3.5-turbo`, `text-embedding-ada-002`, `text-embedding-3-*`
`p50k_base`	Code models, `text-davinci-002`, `text-davinci-003`
`p50k_edit`	Edit models like `text-davinci-edit-001`, `code-davinci-edit-001`
`r50k_base` (or `gpt2`)	GPT-3 models like `davinci`

Context sizes

get_context_size uses the GPT-6 family default for any name beginning with gpt-6, including variants and snapshots. GPT-6 tokenizer lookup and token-budget helpers are unsupported because no encoding mapping is available.

Model	Context window
`gpt-6*`	1,050,000
`gpt-5.4`, `gpt-5.4-pro`	1,050,000
`gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`	1,047,576
`gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5.4-mini`, `gpt-5.4-nano`	400,000
`o1`, `o3`, `o3-mini`, `o3-pro`, `o4-mini`	200,000
`codex-mini`	200,000
`gpt-oss`	131,072
`gpt-4o`, `gpt-4o-mini`	128,000
`o1-mini`, `gpt-5.3-codex-spark`	128,000
`gpt-3.5-turbo`	16,385
`gpt-4`	8,192

See the examples in the repo for use cases. For more context on the different tokenizers, see the OpenAI Cookbook

Encountered any bugs?

If you encounter any bugs or have any suggestions for improvements, please open an issue on the repository.

Acknowledgements

Thanks @spolu for the original code, and .tiktoken files.

License

This project is licensed under the MIT License.

tiktoken-rs