Token Counting Explained: tiktoken, Anthropic, and Gemini (2025 Guide)

Token counting is how providers measure prompt and response size for pricing, context limits, and rate limits. Tokens are not characters or words; they are model-specific units produced by a tokenizer. Understanding how tokenization works—and how models differ—is key to budgeting costs and avoiding truncation.

How Tokenization Works

Most modern models use subword tokenization. Common approaches include Byte Pair Encoding (BPE) and SentencePiece-like unigram models. Tokenizers split text into byte sequences or subwords based on training statistics, so the same sentence may produce different token counts across providers.

OpenAI: tiktoken Overview

OpenAI models are typically measured with tiktoken, an efficient BPE tokenizer. In JavaScript, use js-tiktoken to estimate tokens for prompts and responses.

// TypeScript (Node)
// npm i js-tiktoken
import { encoding_for_model } from "js-tiktoken";

const enc = encoding_for_model("gpt-4o-mini"); // choose a close encoding
const prompt = "Explain zero-copy streaming in Node.js and give examples.";
const ids = enc.encode(prompt);
console.log("tokens:", ids.length);
enc.free();

Note: Pick an encoding close to your target model. Exact counts may vary by release and provider framing (system prompts, tool calls).

Anthropic: Tokenizer and API

Anthropic models (Claude family) use a tokenizer distinct from tiktoken. Anthropic provides an official API and SDKs; for local estimation you can use their tokenizer utilities. See the official guide on token counting for Claude here.

// TypeScript (Node)
// npm i @anthropic-ai/sdk
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY! });
const messages = [
  { role: "user", content: "Contrast streaming vs. non-streaming responses in Claude." },
];

// Official token counting (recommended)
const count = await client.messages.countTokens({
  model: "claude-3-5-sonnet-20240620",
  messages,
});
console.log("anthropic input tokens:", count.input_tokens);

// Normal generation call (actual usage metrics come back on responses)
const msg = await client.messages.create({
  model: "claude-3-5-sonnet-20240620",
  max_tokens: 400,
  messages,
});

Google Gemini: Token Counting API

Gemini exposes a dedicated countTokens endpoint to estimate usage before generating. This is available in both the Generative Language API and Vertex AI. See the API reference at ai.google.dev/api/tokens.

// TypeScript (Node)
// npm i @google/generative-ai
import { GoogleGenerativeAI } from "@google/generative-ai";

const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);
const model = genAI.getGenerativeModel({ model: "gemini-1.5-pro" });

const { totalTokens } = await model.countTokens({
  contents: [{ role: "user", parts: [{ text: "Explain embeddings vs. tokens" }] }],
});
console.log("gemini tokens:", totalTokens);

Mistral: Tokenizer and Counting

Mistral models (e.g., Mistral 7B, Mixtral) use a SentencePiece‑style tokenizer similar to LLaMA families. For local estimates, use the model’s tokenizer from Hugging Face; API responses also include usage fields you can log for ground‑truth counts. See Mistral’s tokenizer guide here.

# Python (Hugging Face Transformers)
# pip install transformers sentencepiece
from transformers import AutoTokenizer

# Choose the exact tokenizer for your target model
# Examples: "mistralai/Mistral-7B-Instruct-v0.2" or "mistralai/Mixtral-8x7B-Instruct-v0.1"
tok = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
text = "Explain rotary position embeddings and KV cache tradeoffs."
ids = tok(text).input_ids
print("mistral tokens:", len(ids))

Tip: Match the tokenizer to the exact checkpoint you are targeting. Small vocabulary changes can shift counts. When validating costs, prefer API usage fields from real calls.

Why Counts Differ Across Models

• Different vocabularies: Token sets vary by provider and model family.
• Pre/post framing: System prompts, tool schemas, and safety wrappers add hidden tokens.
• Multimodal inputs: Images, audio, and structured tool calls have provider‑specific counting rules.

Cost Estimation Tips

See Also

• AI Code Reviews Guide: how token budgets interact with review depth and latency.
• How AI is Transforming Code Completion: practical tips to reduce prompt size while retaining context.