Just in
TechXbox Game Pass November Additions: 3 More Games, 8 Leaving Nov. 30TechUbisoft Motherbrain AI Unit Launches as H1 Net Bookings Fall 16.5%AIChatGPT Space Rolls Out to Plus Users, Replacing the Library TabTechAnthropic IPO Closes With $119.5 Billion in Combined Gross ProceedsAIRiemann Hypothesis Proof Certified by Lean and Mathlib CommunitiesTechCairn AI Agents Found a Zero-Day in MONAI Label, Amazon Says
AI · Guide

What Is a Token in AI? The Numbers Behind What You Pay

A token is the small chunk of text, image or audio an AI model reads and writes, and it is the unit you are billed by. Here are the key numbers, plus a worked cost example.

USA Publishing Group7 min read
Share
Hands holding a tablet showing an AI logo, illustrating how AI models read and bill text in small units called tokens
Photo: Jo Lin / Unsplash

Key takeaways

  • OpenAI and Google both say one token is about 4 characters of English, and OpenAI puts 100 tokens at roughly 75 words.
  • Images, audio and video also become tokens, with Gemini counting 258 tokens for a small image and 32 tokens per second of audio.
  • Output tokens cost more than input tokens, so Claude Sonnet 5.5 charges $2 per million input tokens and $10 per million output tokens at the time of writing.
In this story 9 sections
  1. 4 characters: what a token actually is
  2. 75 words per 100 tokens: turning tokens into something you can picture
  3. 258 tokens per image: pictures and sound count too
  4. 1 million tokens: the context window
  5. $2 and $10: why input and output are priced differently
  6. 30% more: the same text can cost a different number of tokens
  7. 50% off: discounts that change the math
  8. A worked example: summarizing 1,000 reports
  9. Using these numbers before you spend

Four characters. That is roughly how much English text fits in one token, and if you have ever wondered what is a token in AI, that number is the best place to start. A token is the small chunk of text (or image, or audio) that a model reads and writes, and it is also the unit that AI companies count when they send a bill. Here are the numbers worth knowing.

  • About 4 characters of English per token, according to both OpenAI and Google.
  • About 75 words per 100 tokens, per OpenAI. Google puts it at 60 to 80 words.
  • About 100 tokens per paragraph and about 30 tokens for one or two sentences, by OpenAI's rule of thumb.
  • 258 tokens per small image in Gemini, and 32 tokens per second of audio.
  • 1,048,576 tokens is the shared context window for Google's Gemini 4 models.
  • $2 per million input tokens and $10 per million output tokens for Claude Sonnet 5.5 at the time of writing (November 2026).
  • About 30% more tokens for the same text with the newer tokenizer used by Claude 4.7 and later, per Anthropic.
  • 50% off both input and output tokens when you use Anthropic's Batch API.

4 characters: what a token actually is

A large language model does not read words the way you do. Before it sees your prompt, a tool called a tokenizer cuts the text into pieces and turns each piece into a number. Google's documentation says a token can be a single character like "z" or a whole word like "cat," while long words get broken into several tokens.

Think of it like Scrabble tiles that come in different sizes. Common words get one big tile. Rare words, names and typos get spelled out with several smaller tiles. OpenAI notes that even capitalization and spacing matter: " red", " Red" and "Red" can each be a different token.

The 4 character figure is an average, not a rule. Short common words often take one token each. Punctuation, numbers and code tend to eat more.

75 words per 100 tokens: turning tokens into something you can picture

The handiest conversion comes from the OpenAI Help Center: 1 token is about three quarters of a word, so 100 tokens is about 75 words. OpenAI also says 1,500 words comes to roughly 2,048 tokens.

For scale, OpenAI counted the transcript of the US Declaration of Independence at 1,695 tokens. By the same math, a 600 word blog post is about 800 tokens.

Language changes the count. OpenAI's example is the Spanish phrase "Cómo estás," which is 10 characters but 5 tokens. That higher token to character ratio means the same message can cost more in some languages than in English.

258 tokens per image: pictures and sound count too

Tokens are not only text. Google's Gemini token guide lists how other media gets counted:

  • An image no larger than 384 pixels on each side counts as 258 tokens. Bigger images are split into 768 by 768 tiles, and each tile counts as 258 tokens.
  • Audio counts at 32 tokens per second, so a one minute clip is about 1,920 tokens.
  • Video at default resolution runs about 100 tokens per second, or around 6,000 tokens for a minute.

This is why uploading a long video or a stack of photos can burn through a usage limit much faster than typing a question. The same goes for files: when you upload a PDF to ChatGPT, the text inside it becomes tokens the model has to read.

1 million tokens: the context window

The context window is the total number of tokens a model can hold in view at once. Google says its Gemini 4 models share a single window of 1,048,576 tokens across your input, the model's thinking and its output, with a default cap of 65,536 output tokens per reply.

Everything in a conversation fills that space: your messages, the model's answers, any files, system instructions and tool definitions. When a long chat starts "forgetting" early details, the context window is usually the reason. Starting a fresh chat with a short summary often fixes it.

$2 and $10: why input and output are priced differently

When you pay for an AI API, you are billed per token, and output almost always costs more than input. On Anthropic's pricing page at the time of writing, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. Claude Opus 5.5 is $4 and $20. Claude Haiku 5.5 starts at $0.10 and $0.50 for prompts up to 100,000 tokens.

The gap is large: on that table, output costs five times the input rate for both Sonnet 5.5 and Opus 5.5. In practice, a request that asks for a long answer costs far more than one that asks for a short one, even if your prompt is identical.

If you use ChatGPT, Claude or Gemini through a monthly plan, you do not see a per token bill. Tokens still matter, though, because message caps and usage limits are built on them. Our ChatGPT Plus vs Free comparison covers how those plan limits differ.

30% more: the same text can cost a different number of tokens

Every model family has its own tokenizer, so the same paragraph will not produce the same token count everywhere. Anthropic says the tokenizer in Claude 4.7 and later models produces about 30% more tokens for the same text than the one in Sonnet 4.6 and earlier, though the exact increase depends on the content.

That matters when you compare prices. A model with a lower price per million tokens is not automatically cheaper if it turns your text into more tokens. The only fair comparison is to run the same sample through each model and look at the token counts it reports.

50% off: discounts that change the math

Two discounts show up often. Anthropic's Batch API, which processes requests in the background instead of right away, cuts both input and output prices by 50%. Prompt caching lets you reuse a long, repeated prompt (like a big set of instructions) and pay much less for it on later calls. Anthropic lists cache reads at 0.1 times the base input price on most models.

A worked example: summarizing 1,000 reports

Say you run a small business and want Claude Sonnet 5.5 to summarize 1,000 reports, each about 3,000 words, into a 400 word summary.

  1. Input per report. 3,000 words divided by 0.75 words per token is 4,000 tokens. Add about 200 tokens of instructions, for 4,200 input tokens.
  2. Output per report. 400 words divided by 0.75 is about 533 tokens.
  3. Input cost. 4,200 × $2 ÷ 1,000,000 = $0.0084.
  4. Output cost. 533 × $10 ÷ 1,000,000 = $0.00533.
  5. Cost per report. $0.0084 + $0.00533 = about $0.0137.
  6. Cost for 1,000 reports. $0.0137 × 1,000 = about $13.73.

Now adjust for reality. The 0.75 words per token rule comes from OpenAI, and Anthropic says its newer tokenizer creates about 30% more tokens. Multiply by 1.3 and the job comes to roughly $17.85. Run it through the Batch API at half price and it drops to about $8.92.

The output side is only about 39% of the cost here, even though it is a much smaller amount of text. Ask for 800 word summaries instead and the output cost doubles.

Using these numbers before you spend

You do not need to count tokens by hand. OpenAI offers an interactive Tokenizer tool, and the Gemini API has a built in token counter that reports input, output and thinking tokens for each request. Paste a real sample of your text, note the count, then multiply by the current price per million.

The cheapest habit is also the simplest: ask for the length you actually need. Clear, specific prompts, like the ones in our guide on how to write better AI prompts, keep answers tight and token counts low. For more plain English explainers, browse our AI section.