LLM tokens are the small pieces of text that a large language model (LLM) reads and writes.
A token can be a whole word or part of a word. Spaces and punctuation marks can also be tokens. The model never sees words the way you do. If you're asking what are LLM tokens, the practical answer is that they're the unit AI tools use to measure text for pricing and limits on how much they handle.
For a marketer, a token is the unit an AI assistant uses to measure your content. Word counts don't apply here. OpenAI's own rule of thumb puts one token at about 4 characters, so 100 tokens equal roughly 75 words of English. The ratio changes with the language and the model. Text in German or Japanese takes more tokens for the same meaning, as does text full of code. Before you budget a prompt, paste the text into the provider's tokenizer and check the real count.
Tokens decide how much of your page an AI assistant can read and how much it can quote. Assistants that search the web split pages into passages measured in tokens. Then they pick the passages that answer the question best. A fact stated plainly in two sentences fits inside one passage and is easy to cite. The same fact spread across a long, padded section gets split apart or cut off. Brand names matter too. A rare or invented name breaks into several odd fragments, and the model has less experience with those fragments as a sequence. Check that your key claims sit close to the headings they answer.
Tokenization turns raw text into numbers the model can process. Most current models use a method called byte pair encoding (BPE), which learns common character sequences from training data.
A common short word such as "the" is one token, with its leading space included. Longer or rarer words split up. A tokenizer breaks "unbelievable" after "un" and "believ". The final piece is "able". Capital letters and spaces change the split as well, so "Apple" and " apple" can get different IDs. Numbers break into chunks of 1 to 3 digits.
Every model caps how many tokens it can handle at once. Marketers need to know 2 limits.
The context window is the total number of tokens the model can hold in one exchange. That total covers your prompt and any pasted documents. The conversation history and the reply count toward it too. OpenAI's GPT-4o has a 128,000-token context window, which is roughly 300 pages of text. When a chat runs past the window, the oldest material drops out or gets compressed, and the model stops "remembering" your early instructions.
The output limit is the most the model writes in a single reply, and it's much smaller than the context window. GPT-4o stops at 16,384 output tokens. Ask for a 15,000-word report and the reply will end early, mid-sentence. Split long jobs into sections, and check each provider's model page, because these limits change with new releases.
Token counts drive the cost and speed of every AI response. They also shape its quality. API providers charge per token, and output tokens cost more than input tokens. Long prompts slow responses down. Bloated prompts also bury your instructions among filler the model has to weigh. Short, specific prompts give sharper answers for less money. Watch the usage figures in your tool's dashboard to see where the tokens go.
The next step is to run one of your own pages through a free tokenizer and compare the token count with the word count. That one test shows what are LLM tokens in practice and how much of your content an assistant takes in.
Look up more AI search and SEO terms in our glossary.
Related terms
All termsSee how AI talks about your brand