What Is a Token in AI? 2026 Guide With Real Costs

A token is the chunk of text an AI model reads and writes: a word, part of a word or a punctuation mark. OpenAI's rule of thumb is about 4 characters or 0.75 English words per token, and Google's is 100 tokens for 60 to 80 words (verified on vendor docs, October 2026). APIs bill per million tokens, and output usually costs five times input or more.

Here is why that matters at work: the same 1,000-word answer costs less than a tenth of a cent on a budget model and about 7 cents on a flagship one. Multiply by every request your team or your app sends, and the choice of model becomes a budget line.

How many words is a token?

Source: OpenAI key concepts and Google Gemini token docs, read 28 October 2026. Token counts are our arithmetic on each vendor's rule.

So 1,000 tokens is roughly 750 English words by OpenAI's rule, or 600 to 800 by Google's. Both vendors also put it at about 4 characters per token.

These are averages for English, and each model family uses its own tokenizer, so the same text produces different counts on OpenAI, Claude and Gemini models. Short common words like "the" are usually one token; Google's docs note that long words are broken into several, and OpenAI's example splits "tokenization" into "token" and "ization". Code, numbers and other languages often split differently, so measure your own text when the bill matters.

The word also has other meanings. In crypto, a token is a digital asset, and in security, an access token proves who you are. This page is only about tokens in large language models.

How much does a token cost?

APIs price tokens per million, with separate input and output rates. Here are current list prices for a spread of models from three vendors:

Standard API rates, read on each vendor's page on 28 October 2026. OpenAI figures are its standard short-context rates.

The spread is a factor of 100 between the cheapest and the most expensive output price in this table. Batch processing halves the price on OpenAI and Anthropic, and Gemini lists separate, lower batch rates, so work that can wait a few hours costs much less.

Watch for scheduled price changes. Google lists Gemini 4 Argon at double its current rates from 1 December 2026, so a budget set in November will be wrong in December.

What everyday tasks cost in tokens

Dupple data: we priced three common work tasks on five models, using OpenAI's 0.75-words-per-token rule and the list prices above, read on 28 October 2026. The assumptions: an email reply reads 300 words and writes 300 (400 tokens each way); a 1,000-word article comes from a 100-word prompt (133 tokens in, 1,333 out); a 50-page, 25,000-word PDF is summarized into 500 words (33,333 tokens in, 667 out).

All totals are our arithmetic: (input tokens x input price + output tokens x output price) / 1,000,000. Until 30 November, Gemini 4 Argon costs half its column, the same as Sonnet 5.5.

A single task is cheap on any model. The numbers matter at volume: 1,000 articles cost about $0.68 on gpt-6-luna and about $68 on gpt-6-astra, at list price. The PDF row also shows that long inputs dominate the bill on reading-heavy work, while long outputs dominate it on writing-heavy work.

Why do output tokens cost more than input tokens?

On every model in the table above, output costs five times or more what input costs. Generating text is the expensive step: the model produces output one token at a time, while it reads your prompt in one pass.

The practical effect is that a short question with a long answer costs more than a long question with a short answer. If you are paying per token, asking for "a 200-word summary" instead of "a full rewrite" is the single biggest cost lever you control.

Do thinking tokens count?

Yes, and they are billed as output. Reasoning models "think" before answering, and that thinking uses tokens you may never see. OpenAI's reasoning guide says reasoning tokens are not visible through the API but still take space in the context window and are billed as output tokens. Google's price list labels every Gemini output price "including thinking tokens".

This is why a reasoning model can cost far more than its price table suggests. A one-line answer may sit on top of thousands of hidden reasoning tokens. OpenAI warns that a request can even stop before any visible output while you still pay for input and reasoning tokens, so set an output limit and watch the usage figures in the response.

What is a context window in tokens?

A context window is the most tokens a model can handle in one request, counting your prompt and its answer together. OpenAI's docs state the rule directly: for a text generation model, the prompt and the generated output combined must stay within the model's maximum context length.

Context windows are now large. Claude plans list a context window of up to 1M tokens, varying by model, and Google describes Gemini 4 Argon as built for long generation up to 1M tokens. By the 0.75 rule, 1 million tokens is about 750,000 English words (our arithmetic), several long novels.

A big window does not make big prompts free. Every token you send is billed as input on each request, so pasting a whole document into every turn of a chat multiplies the cost. Prompt caching softens this: the Claude, OpenAI and Gemini price lists all show cached input at a fraction of the normal input rate.

How do I count tokens before I send a prompt?

Use the vendor's own counter, because each tokenizer counts differently. OpenAI offers a tokenizer tool linked from its key concepts page. Google's Gemini API has a count_tokens call that returns the input token count before you send the request, and every response reports input, output, thinking and cached tokens.

Images, audio and video count too. Google's token docs say an image up to 384 pixels on both sides counts as 258 tokens, larger images are tiled into 768 x 768 tiles of 258 tokens each, video counts 263 tokens per second and audio 32 tokens per second. A screenshot-heavy prompt can cost more than the text around it.

For a rough budget, the 0.75-words rule is good enough. For anything you will run thousands of times, count a sample with the vendor's tool and use the real number.

Tokens vs messages and credits in chat apps

Chat subscriptions hide tokens behind other units. Claude's paid plans talk about usage per five-hour session, and our Is Claude free? guide explains why Anthropic gives no fixed message count: longer messages and bigger files use more of the allowance. Those factors all come down to how much text the model has to read and write.

If you use AI through a subscription, tokens explain why a long chat with attached PDFs hits the limit sooner than a dozen short questions. If you build on the API, tokens are the bill itself. Our AI chatbot cost guide prices support bots per conversation, including whether building your own on an API is cheaper, and Claude Code pricing works through a subscription vs API comparison for coding.

Related reading

If you route requests across several model vendors, a gateway tracks tokens and spend in one place: see our best LLM gateways guide, or OpenRouter pricing for one popular router. To see where tokens go in production, best LLM observability tools covers monitoring. Running a model on your own hardware removes per-token billing altogether, and best free local LLM tools lists the apps that do it. Claude Opus vs Sonnet shows when a pricier model is worth its tokens, and how to become an AI prompt engineer covers writing prompts that waste fewer of them. DeepSeek adds a twist with time-of-day pricing, explained in Is DeepSeek free?.

Methodology

Rules of thumb come from OpenAI's key concepts page and Google's Gemini token docs. Prices come from OpenAI pricing, Claude pricing and Gemini API pricing, all read on 28 October 2026. Task costs are our arithmetic on the stated assumptions and list prices; real counts vary by tokenizer. We recheck prices every 60 days.

Cite this: Dupple, "What Is a Token in AI? 2026 Guide With Real Costs", November 2026.

FAQ

Price your AI work in tokens, not in prompts: check output length first, count a real sample with the vendor's tool, and watch for scheduled price changes. Start the Dupple X trial →