AI vocabulary: tokens, context, parameters and the 10 words you need to know

Ten words are enough to follow most of what AI apps and AI news are talking about: prompt, model, token, context window, parameters, training data, inference, temperature, hallucination and agent. Each one is explained below in plain language, with a concrete example and what it means for the way you use AI.

How we picked these 10 words

AI jargon runs to hundreds of terms. For this list, we kept the ones that meet three criteria:

• they show up in consumer apps or on their pricing pages;
• they come up again and again in the news about models and assistants;
• they have a real effect on you: on answer quality, limits, price or reliability.

The words follow a logical order, from the most everyday to the more advanced. Each links to its glossary page if you want to dig deeper. For the big picture first, read our guide to what artificial intelligence is.

1. Prompt

A prompt is the text you write to the AI: a question, an instruction, a document to work on.

Example. “Summarise this email” is a prompt. “Summarise this email in three bullet points for a colleague who has not followed the project” is a better one.

Why it matters. The quality of the answer depends heavily on how clear the request is. A good prompt does not have to be long: it gives the context, the goal and the format you expect.

2. Model (LLM)

The model is the computing “brain” that produces the answers. For text, it is called a large language model, or LLM. ChatGPT, Claude and Gemini are assistants; they run on models such as GPT-6 Astra, Claude Opus 5.5 or Gemini 4 Argon.

Example. Within a single assistant, you can sometimes choose between a fast model and a more capable but slower one.

Why it matters. Two assistants can use the same model, and one assistant can switch models. For the full picture, read our explainer on what an LLM is.

3. Token

A token is a small chunk of text: a short word, part of a longer word, or a punctuation mark. It is the unit the model reads and writes.

Example. “The” fits in one token, while a long word like “incomprehensible” is split into several. OpenAI gives a rough rule of thumb for English: one token is about 4 characters, or roughly three quarters of a word.

Why it matters. Length limits and developer pricing are counted in tokens. The site’s token estimator gives you a ballpark figure for any text.

4. Context window

The context window is the maximum amount of text, measured in tokens, that the model can take into account at once: your request, any attached documents, the conversation history and its own reply.

Example. You upload a 200-page report. If it exceeds the window, part of it will be ignored or summarised, often without a clear warning.

Context window (tokens)

• GPT-6 Astra1.05 M
• Claude Opus 5.51 M
• Grok 4.7500 k
• Kimi K2.6262 k
• Qwen3.6 27B262 k

Why it matters. A large window lets you work on long documents. But it is no guarantee: the longer the text, the more likely the model is to miss a detail. It is better to give it the right passages than to give it everything.

5. Parameters

Parameters are the internal numbers a model adjusts during training, like billions of tiny dials. What the model “knows” is stored there.

Example. A model labelled “7B” has around 7 billion parameters. In 2020, OpenAI introduced GPT-3 as a model with 175 billion parameters.

Why it matters. The parameter count tells you about the model’s size and the memory needed to run it, not about its quality. A small, well-trained model can be enough for a simple task.

6. Training data

Training data is all the text, images or audio a model analysed in order to learn: web pages, books, code, licensed content.

Example. A model that read far more English than, say, Welsh or Portuguese may be less comfortable in those languages, or know less about local realities.

Why it matters. This data shapes what the model knows and does not know. It stops at a specific date: without web search, the model is unaware of recent events. Depending on your assistant’s settings, your own conversations may also be used to train a future version.

7. Inference

Inference is the moment when an already trained model produces an answer. Every message you send triggers an inference.

Example. Training a model is like writing a book: long and expensive, but done once. Inference is reading it: it happens again with every question.

Why it matters. Inference determines how fast answers arrive and what each one costs to run. It also explains why free plans cap the number of messages: every reply uses computing power.

8. Temperature

Temperature is a setting that controls how much randomness goes into word choice. At a low setting, the model almost always picks the most likely word; at a high setting, it varies more.

Example. To pull dates out of an invoice, a low temperature gives stable results. To brainstorm names for a community group, a higher temperature gives more variety.

Why it matters. In consumer apps, you do not set it yourself: the provider does. It becomes useful if you work through an API or an automation tool. It also explains why the same question does not always get the same answer.

9. Hallucination

A hallucination is false or invented information that the AI presents with confidence: a date, a quote or a source that does not exist.

Example. You ask for references for an essay; the AI cites a book with a believable title, author and year that was never published.

Why it matters. This is the most important limitation to know about. Check anything precise. Our guide to AI hallucinations sets out six habits to protect yourself.

10. AI agent

An AI agent is a program that uses a model to plan and then act: it chains steps together with tools (browser, email, spreadsheet) until it reaches a goal, with more or less autonomy.

Example. An assistant suggests three trains to Manchester. An agent can compare the times, fill in the booking form and ask you to confirm before paying.

Why it matters. An agent acts on your behalf, so it can also make mistakes on your behalf. Keep control of sensitive steps such as payments or sending messages.

Check yourself You paste a very long contract into an assistant and the answer ignores the last few pages. Which word on this list explains the problem?

The context window. The text probably went beyond the maximum amount the model can read at once. Split the document or provide only the relevant passages.

The recap in one table

Word	In one sentence	Where you see it
Prompt	What you write to the AI	In the text box of any assistant
Model (LLM)	The engine that produces the text	In the model picker
Token	The unit of text read and written	In limits and pricing
Context window	What the model can read at once	In model spec sheets
Parameters	The internal settings it has learned	In model names, such as “7B”
Training data	What the model read to learn	In privacy settings
Inference	Computing an answer	In usage caps and prices
Temperature	The amount of randomness in answers	In developer tools
Hallucination	A mistake stated with confidence	In overly precise answers with no source
AI agent	An AI that acts using tools	In the “agent” modes of assistants

Going further

To see these words at work, read how ChatGPT works, explained simply. And if you come across a term that is not on this list, such as RAG, multimodal or reasoning model, the site’s full AI glossary explains them one by one.