fasttok 0.1.5

High-performance C++ tokenizer and prompt compressor for LLMs — reduce API costs by 10-25%

FastTok

High-performance C++ tokenizer and prompt compressor for LLMs. Reduce OpenAI, Claude, and Gemini API costs by 10–25% with one line of code.

---

Why FastTok?

Every token you send to an LLM API costs money. FastTok's C++ compression engine automatically rewrites your prompts before they're sent — removing verbose phrases and collapsing multi-word expressions — reducing token usage by 10–25% without changing the meaning.

Verified results (tiktoken exact counts):

Text Type	Before	After	Saved
Verbose AI instruction	62 tokens	51 tokens	17.7%
Short tech prompt	27 tokens	23 tokens	14.8%
Business email	58 tokens	54 tokens	6.9%

---

Installation

pip install fasttok

Requires a C++ compiler (MSVC on Windows, GCC/Clang on Linux/macOS).
Pre-built wheels available for Windows, Linux, macOS (Python 3.8–3.12).

---

Quick Start

from fasttok import Tokenizer

# Works with any LLM — auto-detects algorithm
tok = Tokenizer.from_pretrained("gpt-4o")        # OpenAI
tok = Tokenizer.from_pretrained("claude-3-5-sonnet")  # Anthropic
tok = Tokenizer.from_pretrained("gemini-1.5-pro")     # Google
tok = Tokenizer.from_pretrained("llama-3")             # Meta / Groq
tok = Tokenizer.from_pretrained("bert-base-uncased")   # BERT (WordPiece)

text = "For your information, the software development is in order to meet the requirements."

# Compress before sending to API (saves tokens = saves money)
compressed = tok.compress(text)
# "FYI, the software dev is to meet the reqs."

# Count tokens (with optional budget warning)
n = tok.count(text, budget=20)

# Decompress AI response back to human-readable English
readable = tok.decompress("FYI the app dev is ASAP done")
# "for your information the application development is as soon as possible done"

# Trim to fit exactly within a context window
trimmed = tok.smart_trim(text, max_tokens=10)

Real OpenAI Integration

from openai import OpenAI
from fasttok import Tokenizer

tok  = Tokenizer.from_pretrained("gpt-4o-mini")
client = OpenAI(api_key="sk-...")

prompt = "For your information, the application development is approximately done."
compressed = tok.compress(prompt)   # "FYI, the app dev is approximately done."

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": compressed}]
)
print(f"Tokens saved: {tok.count(prompt) - tok.count(compressed)}")

---

Supported Models

Provider	Models	Algorithm
OpenAI	GPT-4, GPT-4o, GPT-3.5, o1, o3	BPE (tiktoken)
Anthropic	Claude 3 / 4 / Sonnet / Haiku	BPE (HF)
Google	Gemini 1.5, Gemma	BPE (HF)
Meta	LLaMA 2 / 3	BPE (HF)
Mistral AI	Mistral, Mixtral	BPE (HF)
Groq	Groq (LLaMA-based)	BPE (HF)
TII	Falcon 40B	BPE (HF)
Alibaba	Qwen 2.5	BPE (HF)
DeepSeek	DeepSeek Coder	BPE (HF)
Google	BERT, DistilBERT, ELECTRA	WordPiece
HuggingFace	RoBERTa, ALBERT	WordPiece

Exact BPE Parity: supply a tokenizer.json or .tiktoken file:

tok = Tokenizer.from_pretrained("llama-3", vocab="path/to/tokenizer.json")

---

Compression Strategies

1. Multi-word → single token: "in order to" → "to", "due to the fact that" → "because"
2. Acronym expansion: "as soon as possible" → "ASAP", "for your information" → "FYI"
3. Technical phrase reduction: "application development" → "app dev"
4. Whitespace/newline normalization: collapse redundant spaces and newlines
5. Bidirectional: decompress() re-expands for human-readable output

---

Performance

• C++17 core with pybind11 — runs GIL-free
• AVX2 SIMD + OpenMP compiled in by default
• Zero network calls at runtime — fully offline after install
• Python 3.8–3.12 supported

---

License

MIT License © 2026 Anique