What Is a Token in AI? The Unit Behind Pricing, Limits, and Forgetting


Read anything about AI models and you'll hit the word token within a paragraph: "supports up to a million tokens of input," "priced per million tokens." Most people mentally translate token as "roughly a word" and move on. That's close enough for casual use — but understanding what tokens actually are explains three things at once: how AI pricing works, why input limits exist, and why a long conversation seems to make the AI forget your earlier instructions.

A token is how AI slices text

AI models don't process text letter by letter or word by word. They break it into tokens — chunks sized for the model to digest. Think of your text as a baguette and tokens as the slices: common words usually fit in one slice, while rare or complex words get split into several. "Cat" is one token; something like "unbelievable" might be split into "un / believ / able."

The slicing rules differ from model to model, so the same sentence can count as a different number of tokens depending on which AI reads it.

Not all languages cost the same

Tokenizers are most efficient on text patterns they saw often in training. Models built primarily on English tend to spend more tokens on the same content written in other languages — Japanese, Korean, Thai and others often consume noticeably more tokens per sentence than English. This has real consequences: the same question can cost more and hit context limits sooner in one language than another. Recent models have narrowed this gap considerably, but it hasn't vanished.

Why pricing is per token

The computational cost of running a model scales with the number of tokens processed, which is why API pricing reads "$X per million tokens." Two details trip people up. First, you pay for both input and output — feeding the model a long report to get a two-line summary costs mostly input tokens. Second, output tokens are typically priced higher than input tokens, so verbose answers cost disproportionately more.

Tokens explain the "forgetting" problem too

A model's maximum working size — its context window — is measured in tokens. When a conversation grows past the window, the oldest content gets pushed out. That's usually what happened when an AI seems to ignore instructions you gave an hour ago: they're no longer in the window at all. The practical fix is to periodically restate the key constraints, or start a fresh conversation with a clean summary.

What token awareness buys you

  • Leaner prompts. Cutting redundant preamble makes responses faster and, on APIs, cheaper.
  • Smarter plan comparisons. Context window size and per-token price are the two numbers that let you compare AI services on substance rather than marketing.
  • Better long-document workflows. Splitting material into sections and summarizing as you go is a token-budget strategy, and it noticeably improves accuracy.

Frequently asked questions

How many words is a token?

There's no fixed conversion. In English, a rough rule of thumb is that a token averages about three-quarters of a word — but it varies by model and by text. For exact counts, use a tokenizer tool.

How can I check the token count of my text?

Most AI providers publish free tokenizer tools where you paste text and see the count. Search for your model's name plus "tokenizer."

Do tokens matter if I just use the chat version?

Pricing-wise, not really — subscriptions are flat-rate. But the context window still applies, so the one thing worth remembering is that very long chats silently drop their oldest content. When quality degrades in a marathon session, that's usually why.

Related on AI Learning Lab: What Is a Context Window? Why AI Forgets · What Are AI Reasoning Models? · What Is a Small Language Model (SLM)?

Comments

Popular posts from this blog

Free vs Paid AI Tools: When Is Upgrading Actually Worth It?

AI Search vs Traditional Search: How to Use Each in 2026

Are AI Certifications Worth It in 2026? A Practical ROI Test