What Are Tokens in AI? Tokens, Limits and Costs Explained

A token is a small chunk of text, usually part of a word, that an AI model reads and writes. In English, one token is about four characters or three quarters of a word, so 100 tokens is roughly 75 words. Tokens are also the meter: they're what API customers pay for, and what fills up a chat's memory and usage limits.

(This guide is about AI language-model tokens, not crypto tokens.)

Most explanations stop there. But the same sentence can be a different number of tokens in ChatGPT, Claude and Gemini; your chat app gets a much smaller memory than the model it runs on; and some of the tokens you pay for are never shown to you. We counted real text ourselves to show how it all works.

#What is a token?

AI models don't read letters or whole words. They read tokens: common words and word pieces from a fixed vocabulary of a few hundred thousand entries. A common word like "the" is one token. A rarer word is split into pieces: "antidisestablishmentarianism" is six.

Small details change the count. Using OpenAI's current tokenizer:

• "hamburger" at the start of a message is 2 tokens ("hamb" + "urger"), but " hamburger" after a space is 1, because the space is part of the token.
• A sentence in ALL CAPS used 20 tokens where the same sentence in normal case used 15, about 33% more.
• "2026" is 2 tokens and "1,234,567" is 6. Numbers, codes and SKUs split into many pieces.

#Why tokens instead of words?

A fixed vocabulary of chunks lets one model handle every language, misspellings, code and made-up words without needing a dictionary entry for each. Common chunks get their own token; rare ones are built from pieces.

#Tokens are the meter

Everything about cost and limits is counted in tokens: what goes in (your prompt, plus any history and instructions), what comes out (the answer, plus any hidden "thinking"), and how much fits in the model's working memory at once.

#How many tokens is a word?

About 1.2 to 1.4 tokens per English word on ChatGPT and Gemini, and roughly 35% more on current Claude models. Each company publishes its own rule of thumb:

#We counted it: one business email

We wrote a normal business email: subject line, greeting, three short paragraphs and a sign-off. 96 words, 558 characters, 119 tokens on OpenAI's tokenizer: 1.24 tokens per word, a little better than the rule of thumb. Gemini counts English at about the same rate. On a current Claude model the same email would be roughly 160 tokens. Anthropic says its newer models use about 30% more tokens than its older ones, which counted close to GPT; independent measurements by Playcode found about 40% more than GPT for English prose, and more for code.

#Other languages cost more, but less than they used to

We translated the same 40-word message into ten languages and counted each with OpenAI's current tokenizer and its older GPT-4-era one:

The "language tax" is real but has shrunk a lot: Hindi went from almost 5× the tokens of English to 1.5×. Pages that say Hindi costs 6 tokens per word are describing older tokenizers. (Our translations are informal; they're fine for comparing counts.)

#Numbers, code and data cost the most

#Quick conversions

• 1,000 tokens ≈ 750 words, about one and a half single-spaced pages.
• 54,000 tokens (ChatGPT Plus's chat window) ≈ 40,000 words, a short novel.
• 1 million tokens ≈ 750,000 words. Google puts it at about 8 average novels or 50,000 lines of code.

#ChatGPT vs Claude vs Gemini: same text, different count

Every company uses its own tokenizer, and they don't agree. That matters more than most people realize, because prices are quoted per token.

In 2026, the three companies converged on identical list prices: Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon (until November 30) all charge $2 per million input tokens and $10 per million output. But if Claude counts roughly 35% more tokens for the same English text, the same work costs roughly 35% more there. Price per token isn't price per task. Our LLM API pricing comparison covers this in detail, and PromptSaver's cost estimates account for it.

Gemini has advantages of its own: independent tests found it uses fewer tokens than GPT for Chinese, and Google counts PDF pages at a flat 258 tokens each (more on that below).

#Why answers cost 5× more than prompts

At OpenAI, Anthropic and Google, mainstream models charge five times more for output tokens than input tokens (six to eight times on some Gemini models). The technical reason, roughly: a model reads your whole prompt at once, in parallel, but writes its answer one token at a time, with each new token depending on the ones before it. Writing keeps expensive hardware busy far longer per token.

#Worked example: one email reply

Ask an AI to reply to that email. The input is 177 tokens (our instruction plus the email); the reply is 140 tokens, about 109 words. The answer is 44% of the tokens but 80% of the bill. At list prices:

The cheapest way to cut that bill isn't a shorter email; it's asking for a shorter reply and choosing the right model.

#Context windows explained: the AI's working memory

A context window is the most text a model can consider at once: your messages, its replies, any files, instructions and hidden thinking, all counted in tokens. Today's top models offer about 1 million tokens through the API (1.05 million for OpenAI's GPT-6 family, 1 million for current Claude models and for Gemini).

#Your chat app gets a fraction of that

The apps most people use give you a much smaller window than the model has, and part of it is taken by the app's own instructions and memories. As listed by each company in October 2026:

Sources: ChatGPT pricing, Google Help, Claude Help. ChatGPT Plus's everyday window is about 5% of what the same models offer developers.

#What happens when you hit the limit

Through the API, a request that's too long fails or the answer gets cut off. In the apps, exactly what happens isn't fully documented; Anthropic says its interfaces can manage the window on a "first in, first out" basis, dropping the oldest messages first. Either way, the AI stops seeing the beginning of your conversation, which is why it seems to "forget" what you said earlier.

#Token limits vs usage limits

Chat plans have two separate limits. The context window caps how long one conversation can be. Usage limits cap how much work you can do per few hours or per week. Long conversations eat into both.

#Why your chat gets slower, pricier and dumber after 50 messages

This is the part almost nobody explains. Every message re-sends the entire conversation. The model has no memory between turns; the app sends your whole chat history each time.

Here's what that looks like for a chat with about 3,000 tokens of app instructions, where each question is 150 tokens and each answer 450:

• Message 1 sends about 3,150 tokens.
• Message 50 sends about 32,550 tokens, roughly 10 times as many, to answer a question of the same size.
• Across all 50 messages, the model reads about 892,500 tokens. Asking the same 50 questions in fresh chats would read about 157,500. The long chat does 5.7 times the work.
• On the API at Claude Sonnet 5.5 prices (counting tokens the GPT way), that's about $2.01 versus $0.54. Caching can bring the long chat's cost down, but the tokens still fill the window.
• With these numbers, ChatGPT Free's 27K window fills around message 40, and Plus's 54K around message 85.

#Slower

Before writing the first word, the model has to read everything you've sent. The longer the conversation, the longer that pause.

#Dumber

Models get worse at using information as the context grows, a problem researchers call context rot. Chroma's study of 18 models from Anthropic, OpenAI, Google and others found performance dropped as input grew, even on simple tasks; a focused 300-token prompt beat a 113,000-token conversation history at answering questions about that history. Anthropic describes models as having a limited "attention budget."

#What to do about it

Start a new chat for each new topic. When a conversation gets long, ask for a five-line summary, open a fresh chat and paste it in. Keep reference files in a Project instead of re-uploading them. (PromptSaver Pro writes these fresh-start summaries for you.)

#Hidden tokens you never typed

#App instructions and memories

Chat apps add their own instructions, your saved memories and custom instructions to every message, inside the same window. Tools and connectors add more; Anthropic notes tools on Claude add a hidden instruction of about 300 tokens to each request.

#Thinking tokens: billed but invisible

Reasoning models "think" before answering, and you pay for every thinking token at the output price, even though you only see a summary or nothing at all. OpenAI says reasoning can run from a few hundred to tens of thousands of tokens per request, and warns you can pay for the thinking and get no answer if the limit is too low. Anthropic and Google bill the full reasoning, not the summary.

Back to our email reply on GPT-6.1 Sol: add 2,000 thinking tokens and it goes from $0.0018 to $0.0218, about 12 times more, for the same visible answer. On a chat plan, extended thinking uses up your limits faster; Anthropic's help center recommends turning it off when you don't need it.

#Images

Images are converted to tokens by size. On current Claude models, a 1000×1000 image is about 1,300 tokens and a 1920×1080 screenshot about 2,700 (Anthropic). OpenAI's GPT-6 Astra counts a 1024×1024 image as about 1,230 tokens. Gemini counts small images as 258 tokens and splits larger ones into tiles of 258 each. Resizing before you upload is free savings.

#PDFs

The companies count documents very differently. Google counts each PDF page as 258 tokens (Gemini docs). Anthropic estimates 1,500 to 3,000 tokens per page for text, and also processes each page as an image (Anthropic). The same 30-page document could be about 7,700 tokens on Gemini or 45,000 to 90,000 on Claude. Send only the pages you need.

#Web search and tools

Search results become input tokens, in that turn and every later one. Claude's API also charges $10 per 1,000 web searches on top.

#How to count tokens

Quick estimate, no tools: divide the number of characters by 4 for ChatGPT or Gemini, and add about 35% for current Claude models. Or multiply English words by 1.3.

Free tools:

• PromptSaver's token counter counts any text for ChatGPT, Claude and Gemini right in your browser, and shows what it costs on every current model.
• PromptSaver's optimizer shows the token count and cost of a prompt as you type, then rewrites it shorter.
• OpenAI's Tokenizer shows how text splits into tokens for its models.
• Anthropic's token counting endpoint is free to use and counts text, images and PDFs for any Claude model (docs).
• Google's countTokens does the same for Gemini (docs).

On your bill: API responses report input, output and reasoning tokens for every request, so you can see exactly where the tokens went.

#7 ways to use fewer tokens

1. Ask for the length and format you need. Output is the expensive part. Our guide to writing better prompts shows how.
2. Start a new chat for each topic, carrying over a short summary.
3. Turn off thinking modes for simple tasks.
4. Paste only the relevant part of a document, and only the pages you need from a PDF.
5. Resize images before uploading.
6. Write numbers and data plainly, and avoid ALL CAPS.
7. Use the smallest model that does the job well. See our guide to reducing AI costs for the full list, with math.

#Frequently asked questions

#What is a token in AI, in simple terms?

A token is a chunk of text, usually a word or part of one, that an AI model reads and writes. In English, a token averages about four characters or three quarters of a word. Models measure prompts, answers, memory and price in tokens.

#How many words is 1,000 tokens?

About 750 English words on ChatGPT, according to OpenAI, and 600 to 800 according to Google's guidance for Gemini. Current Claude models use roughly 35% more tokens for the same text, so 1,000 Claude tokens is closer to 550 words.

#Do ChatGPT, Claude and Gemini count tokens the same way?

No. Each uses its own tokenizer. ChatGPT and Gemini count English text at about the same rate, while current Claude models produce roughly 35% more tokens for the same text. Since prices are per token, the same task can cost more on one than another at the same list price.

#Why do output tokens cost more than input tokens?

Models read your whole prompt at once but write the answer one token at a time, which takes more computing work per token. At OpenAI, Anthropic and Google, output costs about five times input on mainstream models.

#What is a context window?

It's the most text an AI model can consider at once, counted in tokens: your messages, its replies, files, instructions and hidden thinking. Top models offer about 1 million tokens through the API, but chat apps often give you less, such as 54,000 on ChatGPT Plus.

#Why does ChatGPT forget things in a long conversation?

Long chats can outgrow the app's context window, and the app then has to drop or condense older messages. Even before that, models get worse at using information buried in very long conversations. Starting a new chat with a short summary fixes both.

#Do I pay per token on ChatGPT Plus or Claude Pro?

No. Subscriptions charge a flat monthly price with usage limits and a context window. Per-token billing applies to the API. But tokens still matter on a plan: longer prompts, chats and answers use up your limits faster.

#What are reasoning tokens, and do I pay for them?

Reasoning or thinking tokens are the hidden steps a model works through before answering. OpenAI, Anthropic and Google all bill them as output tokens, the expensive kind, even though you see only a summary or nothing. They can multiply a request's cost many times over.

#How many tokens is an image or a PDF page?

On current Claude models, a 1000×1000 image is about 1,300 tokens and a PDF page about 1,500 to 3,000. Gemini counts small images and each PDF page at 258 tokens. OpenAI's GPT-6 Astra counts a 1024×1024 image at about 1,230.

#How we measured

Our counts use OpenAI's current o200k_base tokenizer (and the older cl100k_base for comparison) through the open-source tiktoken library, on text we wrote ourselves, on October 30, 2026. Claude figures are estimates, about 35% above GPT counts, based on Anthropic's published 30% increase over its older models and independent measurements of about 40%; Anthropic's free token counter gives exact counts. Prices are list prices from OpenAI, Anthropic and Google on the same date. App window sizes are as each company lists them and change often.