If you have used ChatGPT, Claude, or any other large language model in your business, you have been billed in tokens, whether you noticed it or not. Tokens are the unit of measurement for everything AI does. They are how the model reads what you write, how it produces its answer, and how every API provider in the world calculates your bill at the end of the month.
Most business owners using AI tools have never been told what a token actually is. That is fine if you only use the consumer ChatGPT subscription. It becomes expensive when you start integrating AI into your operations, your customer support, your content workflows, or any product. This article explains tokens in business terms, not engineering terms, so you can make informed decisions about which model to use, how much it will cost, and where to cut waste.
What Is a Token, Really?
A token is the smallest piece of text that a language model can process at once. It is not a word, and it is not a character. It is something in between. In English, one token averages roughly 4 characters or about 0.75 words. So 1,000 tokens equals approximately 750 words, which is about one single-spaced page of text.
Concrete examples of tokenization:
- The word "hello" is 1 token
- The word "MerchandisePROS" might be 4 to 5 tokens, broken into "Merchandise", "PR", "OS"
- A standard email of 250 words is roughly 330 tokens
- An entire 90-page PDF is approximately 60,000 tokens
- The full text of every Harry Potter book combined is about 1.5 million tokens
Each LLM ships with its own tokenizer, a small program that splits text into tokens before the model processes it. Different tokenizers produce slightly different counts for the same text. OpenAI uses the cl100k or o200k tokenizers. Anthropic uses its own internal tokenizer. The differences are usually small but they exist.
Why Tokens Matter — Cost
Every modern LLM API charges per million tokens, abbreviated MTok. Input tokens, the text you send into the model, are billed at one rate. Output tokens, the text the model generates back to you, are billed at a separate higher rate. Output tokens cost between 2 and 5 times more than input tokens depending on the provider. The industry-median ratio in 2026 is approximately 4 times output to input.
Here are concrete prices from the major providers as of 2026, listed as input dollars / output dollars per million tokens:
- OpenAI GPT-4.1 Nano — $0.10 / $0.40 per MTok (cheapest tier, for routine work)
- OpenAI o1 reasoning — $15.00 / $60.00 per MTok (frontier reasoning, expensive)
- Anthropic Claude Opus — $5.00 / $25.00 per MTok (frontier, balanced)
- Anthropic Claude Sonnet — $3.00 / $15.00 per MTok (production workhorse)
- Google Gemini 2.5 Flash — $0.15 / $0.60 per MTok (cheap, fast, very large context)
- Google Gemini 2.5 Pro — $1.25 / $5.00 per MTok (mid-tier)
- DeepSeek V3.2 — $0.14 / $0.28 per MTok (open-weight Chinese model, extremely cheap)
The spread between the cheapest and the most expensive frontier model is roughly 100 times. Picking the right model for the job is the single biggest lever a business has on AI cost.
Input vs Output — Why the Gap?
The reason output tokens cost 2 to 5 times more than input tokens has to do with how transformer models actually work. When you send a prompt to a model, the entire prompt is processed in a single forward pass through the neural network. The model reads it once, in parallel, and that step is relatively cheap.
Generating output works differently. The model produces one token at a time, autoregressively. To generate token number 100, the model must run a full forward pass over all 99 tokens it has already generated, plus the original prompt. Generating token number 200 means running a forward pass over 199 tokens. This compounds. A 1,000-token answer requires roughly 1,000 forward passes through the model, while the same 1,000 tokens of input would require only one. That is why output is more expensive.
Context Window — How Much the Model Can "Remember"
The context window is the maximum number of tokens, combining both input and output, that a model can handle in a single request. If you exceed the context window, the request fails or earlier content gets dropped silently. Context windows have grown dramatically in 2026:
- Claude Opus 4.6 and Sonnet 4.6 include 1 million tokens of context at standard pricing
- OpenAI GPT-4.1 supports up to 1,050,000 tokens
- Google Gemini 2.5 Pro includes 1 million tokens with optional 2 million tokens at higher tiers
- Older GPT-4 and GPT-3.5 capped at 8,000 to 128,000 tokens
For perspective, a 1 million token context can hold every email you have written in the last three years, plus a 500-page legal contract, plus a year of internal Slack messages, all in a single request. The implication for business: workflows that previously required complex retrieval systems can now fit into a single prompt.
Three Top LLMs at a Glance
Anthropic Claude — Frontier reasoning, 200K to 1M context, strongest model in coding and long-document reasoning. Used as the engine behind Cursor, Replit, and GitHub Copilot. Claude users skew technical and enterprise.
OpenAI ChatGPT (GPT-4.1, o1) — Largest market share by a wide margin, broadest ecosystem, default consumer brand for AI. Strong across the board, with the o1 reasoning model leading on complex math and science benchmarks.
Perplexity — Different category. It does not train its own frontier model. Instead, it routes queries to Claude or GPT and adds live web search with citations. 170 million monthly visits and the fastest-growing search-replacement product in the AI space.
What This Means for Your Business
Tokens are not an abstraction. They are the unit on your invoice. A business that ignores token math will overpay by 5 to 10 times for the same output. A business that understands tokens can run the same workload at frontier quality for a fraction of what its competitors pay.
Three cost-optimization moves every business should make:
- Cache long system prompts. Both Anthropic and OpenAI offer prompt caching that drops cached input tokens to roughly 10 percent of normal price. If your app sends the same 5,000-token system prompt on every request, caching it saves about 90 percent of input cost.
- Batch non-real-time tasks. Anthropic and OpenAI both offer batch APIs at 50 percent off. If your work does not need an answer in seconds, batching it cuts the bill in half.
- Use cheaper models for routine work. GPT-4.1 Nano at $0.10 / $0.40 handles classification, summarization, and basic writing as well as Claude Opus does, at roughly one fiftieth the price. Reserve the frontier models for jobs that actually need them.
Frequently Asked Questions
How many tokens is one page of text?
Roughly 1,000 tokens equals about 750 English words, which is approximately one single-spaced page of text. The exact ratio depends on the tokenizer used by the model and the language. English averages about four characters per token. Spanish, Portuguese, and other Romance languages tokenize around 10 to 20 percent larger than English because of accented characters and longer average word length.
Why do output tokens cost more than input tokens?
Input tokens are processed once in a single forward pass through the model. Output tokens are generated autoregressively, meaning the model performs a full forward pass for every single output token it produces. That additional compute is why output is billed 2 to 5 times more than input across every major LLM provider. The industry-median ratio in 2026 is approximately 4 times output to input.
What is a context window?
A context window is the maximum number of tokens, combining both input and output, that a model can process in a single request. Claude Opus 4.6 and Sonnet 4.6 include a 1 million token context at standard pricing. OpenAI GPT-4.1 supports up to 1,050,000 tokens. Google Gemini 2.5 includes 1 million tokens. Anything beyond the window must be summarized, retrieved selectively, or truncated.
Can I reduce token costs?
Yes. Three proven techniques: cache long, repeated system prompts, which drops cached input tokens to roughly 10 percent of normal price on Anthropic and OpenAI. Use the batch API for non-real-time tasks for a 50 percent discount on both Anthropic and OpenAI. Choose smaller, cheaper models like GPT-4.1 Nano or Gemini 2.5 Flash for routine tasks and reserve the frontier models for jobs that actually need them.
Do tokens differ between English and Spanish?
Yes. Spanish text tokenizes roughly 10 to 20 percent larger than the equivalent English text. Accented characters such as a, e, i, o, u with tildes often consume an extra token, and Spanish words are on average longer than English words. The same paragraph translated from English to Spanish will cost slightly more in API tokens. Plan for that when budgeting bilingual workloads.
"If you do not understand tokens, you do not understand what AI is actually charging you for. Every business owner using ChatGPT or Claude should know how to read a token bill."
- Diego Medina F, Founder of MerchandisePROS