Explainers

AI Jargon, Decoded: Tokens, Context Windows, Open Weights and More

The plain-English glossary you need to understand AI news, model launches and our weekly rankings.

By Alex · · Updated · 6 min read

In this guide
  1. Model
  2. Token
  3. Input vs. output tokens
  4. Context window
  5. Benchmark
  6. Arena / Elo score
  7. Open weights
  8. Reasoning / effort levels
  9. Agent
  10. Multimodal / Omni
  11. Diffusion model
  12. Hallucination
  13. Prompt and system prompt
  14. Training data and knowledge cutoff
  15. Parameters
  16. Fine-tuning
  17. RAG (retrieval-augmented generation)
  18. Embeddings
  19. API
  20. Inference
  21. Quantization and distillation
  22. Temperature
  23. Guardrails
  24. Flash / Mini / Luna / Haiku
  25. Pro / Opus / Astra / Fable
  26. Quick reference
  27. Common misunderstandings

AI news is full of words nobody explains. This is the glossary we wish we'd had, and it covers every term you'll see in our weekly standings.

Model

The "brain" behind an AI app. ChatGPT is an app, GPT-6 Sol is a model. One app can switch between several models, and the same model can power many different apps.

Token

The unit AI models read and write in, roughly three-quarters of a word. "Hamburger" might be two or three tokens. AI companies charge developers per million tokens, which is why you'll see prices like "$2 in / $10 out".

Input vs. output tokens

Input is what you send (your question plus any documents). Output is what the AI writes back. Output usually costs more, often 4–5x more. Try it in our cost calculator.

Context window

How much the model can "hold in mind" at once, measured in tokens. A bigger context window means you can give it a whole book or a big codebase in one go.

Benchmark

A standardised test for AI. Examples include Terminal-Bench (can it get real tasks done in a computer terminal?) and AutomationBench (can it automate multi-step work?). Useful, but labs pick which ones to publish, so read them with a pinch of salt.

Arena / Elo score

LMArena shows people two anonymous AI answers side by side and asks which is better. Millions of votes produce an Elo score, the same kind of rating chess uses. A difference of a few points is basically a tie.

Open weights

A model whose "brain file" you can download and run yourself, on your own computer or server. That's good for privacy and customisation. It isn't always "open source" in the strict sense, because the license can still restrict how you use it. See our Open-Weight Watch.

Reasoning / effort levels

Many 2026 models can "think longer" before answering. Higher effort gives better results on hard problems, but it's slower and uses more tokens, so it costs more.

Agent

An AI that doesn't just answer, but takes actions: browsing, clicking, running code, sending emails, often over many steps. Agentic ability is one of the biggest differences between the top models this year.

Multimodal / Omni

A model that handles more than text, such as images, audio and video, as input, output or both. Google's Gemini Omni models currently top the video leaderboard.

Diffusion model

A different way of generating content. Instead of writing one word after another, it refines everything in parallel. It's traditionally used for images and video, and now also for fast text models like Mercury 2.5.

Hallucination

When an AI confidently states something false. Newer models hallucinate less (OpenAI says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol), but it still happens. Always check important facts.

Prompt and system prompt

A prompt is what you type to the AI. A system prompt is a hidden set of instructions an app gives the model before your conversation starts, such as "You are a helpful customer service assistant for a bank." It's why the same model can behave very differently in different apps. Our guide to writing better prompts covers the part you control.

Training data and knowledge cutoff

Models learn from a huge collection of text, images and code called training data. The knowledge cutoff is the point where that data stops. Ask about anything after it, and the model either doesn't know or has to search the web. That's why the same question can get an outdated answer from one app and a current one from another.

Parameters

The internal numbers a model learns during training, often counted in the billions or trillions. More parameters roughly means more capacity, but it isn't a reliable measure of quality on its own. Many smaller 2026 models beat much larger older ones, and many labs no longer publish parameter counts at all.

Fine-tuning

Extra training on a smaller, specialised set of examples to adjust a model for a specific job, such as legal writing, a company's support tone or a particular programming language.

RAG (retrieval-augmented generation)

A technique where an app looks up relevant documents first and hands them to the model along with your question. It's how many company chatbots answer questions about their own policies without retraining the model, and it usually cuts down on made-up answers.

Embeddings

A way of turning text into lists of numbers so a computer can measure how similar two pieces of text are in meaning. Embeddings power semantic search ("find documents about this idea, not just these words") and most RAG systems.

API

An application programming interface: the way developers plug a model into their own apps and pay per token rather than a monthly subscription. When you see prices like "$2 in / $10 out per million tokens," that's API pricing.

Inference

The act of running a model to get an answer, as opposed to training it. When labs talk about "inference costs," they mean what it costs to serve each answer.

Quantization and distillation

Two ways to make models smaller and cheaper to run. Quantization stores the model's numbers with less precision, so it fits on smaller hardware with a small quality trade-off. Distillation trains a small "student" model to imitate a bigger "teacher" model. Many of the fast, cheap models in our value rankings use these tricks.

Temperature

A setting that controls how random a model's answers are. Low temperature gives more predictable, focused answers; high temperature gives more varied, creative ones. Most chat apps choose it for you; developers can set it through the API.

Guardrails

The rules and filters that stop a model from producing harmful content or taking risky actions. Every major lab builds them in, and stricter or looser guardrails are a big part of why models feel different to use.

Flash / Mini / Luna / Haiku

Different labs' names for their smaller, faster, cheaper models. They're often good enough for everyday tasks at a fraction of the price.

Pro / Opus / Astra / Fable

The labs' names for their biggest, smartest, most expensive models.

Quick reference

TermOne-line meaning
TokenAbout three-quarters of a word; the unit AI is billed in
Context windowHow much text the model can consider at once
Knowledge cutoffThe date the model's training data ends
HallucinationA confident but false answer
AgentAI that takes actions, not just answers
Open weightsA model you can download and run yourself
RAGLooking up documents before answering
Fine-tuningExtra training for a specific job
InferenceRunning a model to get an answer
GuardrailsSafety rules built into a model or app

Common misunderstandings

  • "Bigger context window means it remembers me." No. The context window covers one conversation. Separate "memory" features are what carry details between chats.
  • "Open weights means free and unrestricted." Not always. Check the license.
  • "A higher benchmark score means it's better for me." Only if the benchmark matches your task. A coding benchmark tells you little about email writing.
  • "The AI looked it up." Only if it has web search turned on and shows you sources. Otherwise it's answering from training data. See how to fact-check AI answers.

Now you speak fluent AI. Go see who's on top this week →

Keep reading