Understanding LLM Tokens and API Pricing

Published: October 7, 2026 · 8 min read · By JSON Suite Team

If you're using AI APIs like OpenAI's GPT-4, Anthropic's Claude, or Google's Gemini, you've seen pricing in "tokens." But what exactly is a token, and how do you estimate costs? This guide explains everything you need to know.

What is a Token?

A token is the basic unit that language models process. It's not a word, character, or sentence — it's something in between. Tokens are pieces of text that the model's tokenizer splits input into.

Roughly speaking:

1 token ≈ 4 characters (in English)
1 token ≈ 0.75 words (common estimate)
100 tokens ≈ 75 words

How Tokenization Works

Most modern LLMs use subword tokenization (like BPE — Byte Pair Encoding). Common words become single tokens, while rare words are split into pieces:

// Common words = 1 token each "the", "and", "is", "hello", "world" // Less common = split into multiple tokens "tokenization" → ["token", "ization"] // 2 tokens "JSON" → ["JS", "ON"] // 2 tokens "antidisestablishmentarianism" → 6+ tokens // Numbers vary "2024" → 1 token "3.14159" → 3 tokens (often split at decimals) // Non-English uses more tokens "こんにちは" (Japanese "hello") → 3-5 tokens "你好" (Chinese "hello") → 2-3 tokens

Why Tokens Matter

Tokens affect two things:

Cost — APIs charge per token (usually per 1,000 tokens)
Context limits — Models have maximum token limits (context window)

Input vs Output Tokens

Most APIs charge differently for:

Input tokens — Your prompt, system message, and any context you send
Output tokens — The model's generated response

Output tokens are typically 2-4x more expensive than input tokens because generation is more compute-intensive.

Model Pricing Comparison (2026)

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context Window
GPT-4 Turbo	$10	$30	128K
GPT-4o	$2.50	$10	128K
GPT-3.5 Turbo	$0.50	$1.50	16K
Claude 3 Opus	$15	$75	200K
Claude 3 Sonnet	$3	$15	200K
Claude 3 Haiku	$0.25	$1.25	200K
Gemini 1.5 Pro	$3.50	$10.50	1M
Gemini 1.5 Flash	$0.075	$0.30	1M

Note: Prices change frequently. Always check official pricing pages.

Calculating Costs

Here's how to estimate API costs:

// Example: GPT-4o // Input: 2000 tokens, Output: 500 tokens Input cost: 2000 ÷ 1,000,000 × $2.50 = $0.005 Output cost: 500 ÷ 1,000,000 × $10 = $0.005 Total: $0.01 per request // 10,000 similar requests = $100

Context Window Explained

The context window is the total tokens the model can handle in one request (input + output combined). This affects what you can do:

4K tokens — Short conversations, simple tasks
16K-32K tokens — Document summarization, code review
128K tokens — Analyze entire codebases, long documents
1M+ tokens — Process books, massive datasets (Gemini 1.5)

Tips to Reduce Token Usage

1. Be Concise in Prompts

// Verbose (50+ tokens): "I would like you to please help me by summarizing the following text that I have provided below in a way that captures the main points and key ideas." // Concise (10 tokens): "Summarize this text in 3 bullet points:"

2. Limit Output Length

// Use max_tokens parameter { "model": "gpt-4", "max_tokens": 150, // Cap response length "messages": [...] } // Or instruct in prompt "Answer in 2-3 sentences maximum."

3. Use Cheaper Models for Simple Tasks

Not every task needs GPT-4. Use GPT-3.5 or Claude Haiku for:

Simple classification
Text extraction
Basic formatting
Translations

4. Compress Context

// Instead of full conversation history: [30 previous messages = 5000 tokens] // Summarize periodically: "Previous context: User asked about Python, we discussed lists and dictionaries. Now continuing..." [500 tokens saved!]

5. Remove Unnecessary Data

// Don't send full JSON if you only need some fields // Before (200 tokens): { "user": { "id": 1, "name": "John", "email": "...", "address": {...}, "preferences": {...}, ... } } // After (20 tokens): { "name": "John", "email": "john@example.com" }

JSON and Tokens

JSON is commonly used in LLM APIs for structured input/output. However, JSON uses more tokens than plain text:

// JSON format (more tokens due to punctuation): {"name": "John", "age": 30, "city": "NYC"} // ~15 tokens // Plain text alternative: "name=John, age=30, city=NYC" // ~10 tokens // But JSON is worth it for: - Structured extraction - Function calling - Consistent parsing

Different Tokenizers

Each model family uses different tokenizers, so the same text produces different token counts:

Text	GPT-4 (cl100k)	Claude	Gemini
"Hello, world!"	4 tokens	4 tokens	3 tokens
"JSON formatting"	3 tokens	3 tokens	2 tokens
1000-word essay	~1200 tokens	~1100 tokens	~1000 tokens

This is why counting tokens accurately requires using the right tokenizer for your model.

Batch Processing for Savings

Some providers offer batch/async APIs at lower costs:

OpenAI Batch API — 50% discount, 24-hour processing
Anthropic Batches — Similar discounts for async workloads

Great for non-urgent tasks like content generation, data processing, or analysis.

Count Your Tokens

Before sending expensive API calls, count your tokens! Use our free LLM Token Counter to:

Count tokens for GPT-4, Claude, Gemini, and more
Calculate estimated API costs
Optimize prompts before sending
Stay within context limits

It's especially useful when working with JSON APIs where structured data can quickly add up.