Under a video about building an app with an AI agent, a reader asks: "What does this cost as far as tokens you use to build this?" (31 likes). Nearby comments ask the same thing from other angles. "Please tell us how to save tokens. It's amazing but burns token like crazy!" Another: "How many seconds do your tokens last for with all those agents working?" Another, from someone new to agents: "How much token would that cost? Are we talking about spending at least 300usd per month?"
All of those questions count in the same unit. A token is the chunk of text AI models read and write in, and it's also the unit that prices, usage limits, and "chat is full" messages count in. Once the unit makes sense, you can start estimating the answers. The nightly comment scrape this Lab runs read 4,933 comments as of tonight, and 107 of them mention tokens.
What a token is
A token is a piece of text a model processes. Before a model reads your message, the text gets split into tokens: short, common words can be one token, longer words split into pieces, and punctuation and spaces count too. A token can be a single character, part of a word, or a whole word.
The estimates OpenAI publishes for English: one token is about four characters, or about three-quarters of a word, so 100 tokens come out to roughly 75 words. The counts are estimates, not math you can run in your head: the same text can produce different token counts on different models, and other languages split differently than English.
You can watch the split yourself. OpenAI's Tokenizer (platform.openai.com/tokenizer) is free and needs no account: paste in a sentence and it highlights each token and shows a count. A sentence like "I don't think this word counts" is a good one to try, because the apostrophe, the contraction, and the punctuation each change the split. Ten seconds in that tool teaches the concept faster than any definition.
The three places you meet tokens
Prices. When you buy model access through an API, the price list quotes dollars per million tokens, with input and output priced separately. OpenAI's current list charges $10 per million input tokens and $50 per million output tokens on its most expensive flagship model, and 10 cents and 50 cents on its cheapest. That is a hundredfold spread between models that can describe themselves in similar words, and it's why "which model" is a cost question as much as a quality one. What Does It Cost to Run an AI Agent? Subscriptions vs API walks the full arithmetic.
Limits. Chat apps count usage against an allowance instead of a meter. On ChatGPT's free tier, everyday text chats are unlimited within abuse-prevention limits, but tools like image generation, file analysis, and voice carry separate limits; reach one and the app tells you and shows when it resets. Paid plans add higher limits on the same shape: an allowance that refills, rather than a per-token bill.
The context window. A model can only work with so many tokens at once, and a chat that outgrows it shows up as a "chat is full" message. Your messages and the model's replies both take up space, and a few long exchanges can fill the window faster than many short ones. What Is Context in AI? Why It Makes or Breaks Your Results covers the window itself; the token is the unit it fills with.
Why tokens run out fast
Three mechanics explain most fast token use.
Long chats re-read themselves. Each new reply processes the conversation so far, so the same question costs more tokens to ask at the end of a long chat than at the start. Usage grows with the history, even when your new message is short.
Thinking counts as output. Reasoning models produce internal reasoning tokens before the visible answer, and those bill as output whether or not you ever see them. A short-looking answer can carry a much larger token count than its text suggests.
Agents multiply the reads. An agent step reads tool output, page text, or file contents, then often runs again with all of it attached. Commenters under the agent videos notice: "so you'd be burning grok & claude tokens at the same time?" is one reply, and "do these dashboards use additional tokens to run?" is another. The second question has a fairly general answer: most of what an agent reads or produces counts toward usage, and dashboards and logs usually draw from the same allowance.
Making an allowance last
Start a fresh chat when the job changes. History rides along in a long chat, and a job that shares nothing with the previous one pays to carry it anyway.
Paste the material the job needs, and skip the rest. The model can't un-read a paste, so a 40-page document attached to a two-line question spends tokens on 39 pages of it.
Match the model to the job. Routine text work on a cheap model costs a fraction of what the same work costs on a flagship, at the hundredfold spread from the price list above.
Read the limit notices when they appear. The app shows reset timing when you hit one, which turns "I'm locked out" into "this refills at 3 pm."
One honest caveat: you never need to count tokens by hand. The tools meter themselves, and the counts differ by model and language anyway. The skill is knowing what makes the number move, not tracking the number.
Frequently asked questions
How much text is one token?
For English, OpenAI's estimates are one token per four characters, one token per three-quarters of a word, or 75 words per 100 tokens. These are approximations: the same text tokenizes differently across models and languages.
Why did my ChatGPT limit run out so fast?
Tool use draws on the allowance faster than plain chat. Image generation, file and data analysis, voice, and agent-style features carry their own limits or consume the shared allowance at a higher rate, and a reasoning model can bill thinking tokens as output. The app shows the reset time when you hit a limit.
Do tokens cost extra on a paid subscription?
No. Subscription plans include an allowance that refills, and hitting it means waiting for a reset rather than paying per token. Per-token billing is the API's model, where developers pay by the million for input and output separately.
How do I check my token usage?
In a chat app, watch the limit notices: they show when the allowance resets. On the API side, the usage dashboard reports token counts per request and over time.
Do all AI tools use the same tokens?
No. Each model family has its own tokenizer, so the same sentence can produce different counts in different tools, and punctuation, spacing, and capitalization all affect the split. That's why estimates are the standard, not exact counts.
Want to go further?
If the price side is what brought you here, What Does It Cost to Run an AI Agent? Subscriptions vs API prices a job end to end, and ChatGPT Free vs Paid: Is the Upgrade Worth It? sorts out when an allowance bump earns its fee. And if you only take one action from this post: paste a sentence into the Tokenizer and watch it split. It's easier to reason about the unit after you've watched it take apart a sentence you wrote.