AI glossary: 50 terms in plain English

Definitions are ours; the figures inside them come from vendor documentation checked in September 2026. Some vendor links on this site are affiliate links; see our disclosure.

How to use this page

Read it straight through once, about fifteen minutes, or search the page for the word that stopped you. Terms are grouped by nothing except the alphabet, because the same word turns up on a chatbot's pricing page, a developer's API doc and a news story about deepfakes. Where a term is a thing you can buy or switch on, the link goes to the tool card or guide that covers it.

The terms

AGI

Artificial general intelligence: a system that can do most economically useful cognitive work at least as well as a person, rather than one narrow task. There is no agreed test for it, which is why every lab defines it slightly differently and why claims that a model "is AGI" should be read as marketing until someone names the test. As of September 2026 no vendor's pricing page sells it; they sell frontier models.

Agent

A model set up to take actions, not just write text: browsing a site, running code, sending an email, calling other tools, and looping until a task is done. ChatGPT's agent mode, Claude Code, Perplexity Computer and Gemini Spark are all agents. The practical difference from a chatbot is that an agent can spend your money and change your files, so the permissions you grant it matter more than the model behind it. See the chatbot comparison for which plans include one.

Alignment

The research problem of making a model reliably do what its operators and users intend, including refusing harmful requests and not pursuing goals of its own. It covers training methods such as RLHF, the rules written into a system prompt, and testing before release. When a lab says a model was "trained to be helpful, honest and harmless," that is an alignment claim, and a model card is where they are supposed to show their evidence.

API

Application programming interface: the way software, rather than a person in a chat window, sends requests to a model and gets answers back. API access is billed per token and is separate from a chat subscription; paying for ChatGPT Plus or Claude Pro does not include a single API token. The model tracker lists API prices for every current model.

Benchmark

A fixed set of test questions used to score models against each other, such as SWE-Bench for coding or GPQA for graduate-level science. Benchmarks are useful for spotting big gaps and useless for small ones, because labs train with the tests in mind and report their own scores. Treat a vendor's benchmark table as a claim, and treat a five-prompt test on your own work as the evidence; the choosing guide shows how to run one.

Chain of thought

The technique of having a model write out intermediate reasoning steps before its final answer, which measurably improves accuracy on math, logic and multi-step tasks. Early on you triggered it by adding "think step by step" to a prompt; reasoning models now do it automatically and hide or summarize the steps. The steps are billed as output tokens even when you never see them.

Compute

The raw processing capacity used to train and run models, usually measured in GPU-hours or in floating-point operations. Compute is the main cost behind every price on this site: training a frontier model takes months on tens of thousands of chips, and every answer you get costs a slice of a chip's time. When a lab says it has started "its most ambitious pre-training run yet," that is a compute statement.

Context window

The maximum amount of text, in tokens, that a model can hold in one request, counting your prompt, any files, the conversation so far and its own answer. Current flagships from OpenAI, Anthropic, Google and others offer around a million tokens on the API, but chat apps often expose less: the free ChatGPT tier holds 27K tokens for instant answers. When a model "forgets" the start of a long chat, the context window overflowed. Compare windows in the model tracker.

Data training opt-out

The setting that stops a vendor using your conversations to train future models. As of September 2026 it is off by default, meaning training is on, for personal plans at ChatGPT, Claude, Gemini and Perplexity, and each buries it in a different place: Data Controls, Privacy, Gemini Apps Activity and AI data retention respectively. Business and API accounts do not train by default. Flip it on day one; the choosing guide makes it step six.

Deepfake

Synthetic audio, image or video that shows a real person saying or doing something they did not. Voice cloning and image models have made convincing ones cheap, which is why ElevenLabs and similar services require consent checks for cloning a voice and why watermarking has become a policy issue. The defense is procedural, not technical: verify unexpected requests through a second channel.

Diffusion model

The architecture behind most image and video generators. It learns to turn random noise into a picture in many small denoising steps, guided by your text prompt. Nano Banana, Sora and the image tools inside Canva AI are diffusion-based or descended from it. Diffusion models are priced per image or per second of video rather than per token.

Distillation

Training a smaller, cheaper model to imitate the outputs of a larger one, so it inherits much of the quality at a fraction of the running cost. Meta's Llama 4 Scout and Maverick were distilled from a larger teacher model, and most vendors' "mini" and "flash" tiers are distilled from their flagships. It is also why labs restrict bulk API use by competitors: distillation works on rivals' models too.

Edge AI

Running a model on a device near the user, such as a phone, laptop, car or factory sensor, rather than in a data center. The appeal is speed, privacy and working offline; the cost is that only small, often quantized, models fit. Meta's Muse Glimmer and Mistral Small 4 are built for this. See also on-device.

Embeddings

A way of turning a piece of text, an image or a sound into a long list of numbers that captures its meaning, so that similar things end up with similar numbers. Embeddings are how search finds "car insurance quote" when you typed "how much to insure a vehicle," and they are the first step of RAG. Vendors sell embedding models separately and cheaply per token.

Few-shot

Giving the model a handful of worked examples inside the prompt before asking it to do the real task, so it copies the pattern. Two or three examples of emails in your voice, followed by "now write this one," is few-shot prompting, and it is the single most effective trick in our email prompts. Contrast with zero-shot.

Fine-tuning

Continuing to train an existing model on your own examples so it permanently adopts a style, format or domain, rather than being told each time in the prompt. It is a developer feature, priced per training token and per use, and most people who think they need it actually need few-shot prompting or RAG. Some frontier models, including GPT-6 Astra, do not offer it at all.

Frontier model

A model at or near the best available capability at the time, typically the flagship of a major lab. In September 2026 that means GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5, with Gemini 3.1 Pro, Muse Spark 1.3, Grok 4.7 and DeepSeek V4 Pro close behind. The term also carries a regulatory meaning: frontier models get extra safety testing before release, which is why OpenAI delayed Astra to strengthen cybersecurity safeguards.

GPU

Graphics processing unit: the chip that does the arithmetic behind training and inference, originally designed for video games and now the scarcest resource in the industry. Nvidia's data-center GPUs dominate; the H100 was the reference chip for running a model like Llama 4 Scout. When a vendor's service slows at peak hours, it is short of GPUs.

Guardrails

Rules and filters around a model that block certain inputs and outputs regardless of what the model would otherwise say: refusing to help with weapons, screening for personal data, stopping an agent from spending above a limit. Guardrails sit outside the model, in the product, which is why the same model can behave differently in two apps. Grok 4.7 shipped with a rebuilt safeguard stack; that is guardrails.

Hallucination

A confident, fluent, wrong answer: an invented citation, a made-up price, a function that does not exist. Every current model does it, and the tone gives no warning. The defense is to ask for sources and check them, to prefer tools that show citations like Perplexity, and to include a checkable-fact prompt in any trial. Reasoning models hallucinate less on hard questions and still hallucinate.

Inference

Running a trained model to produce an answer, as opposed to training it. Every chat message, API call and image generation is inference, and it is what the per-token prices pay for. Vendors talk about "inference costs" when they cut prices or add off-peak rates, as DeepSeek did in August 2026.

Latency

How long you wait for a response, from the first token to the last. Reasoning models trade latency for accuracy: an answer at high effort can take a minute where a flash-class model answers in two seconds. Anthropic labels its lineup from "fastest" to "slower" for this reason. For voice and live translation, latency is the whole product.

LLM

Large language model: a neural network trained on enormous amounts of text to predict the next token, which turns out to be enough to write, summarize, translate, code and reason. ChatGPT, Claude, Gemini and Grok are chat products built on LLMs. "Large" is relative; today's small models are bigger than the first ones that earned the name.

MCP

Model Context Protocol: an open standard, described by its maintainers as a USB-C port for AI applications, that lets an assistant connect to outside tools and data such as a calendar, a database or a design file in one consistent way. Claude, ChatGPT, Cursor and Gemini all support it, so a connector built once works across them. When a chatbot's pricing page lists "connectors," MCP is usually what is underneath.

Mixture of experts

An architecture where a model contains many specialist sub-networks and activates only a few of them for each token, so a model with a huge total parameter count runs at the cost of a much smaller one. DeepSeek V4 Pro has 1.6 trillion parameters in total but uses 49 billion per token; Mistral Large 3 is 675 billion total and 41 billion active. It is why open-weight models got both bigger and cheaper in 2026.

Model card

The document a lab publishes with a model describing what it is for, what data it was trained on in general terms, how it scored on safety and capability tests, and what it should not be used for. Google DeepMind and Anthropic publish them for every release; a vendor with no model card is asking you to trust a black box. Reading one is the fastest way to find a model's context window and knowledge cutoff.

Multimodal

A model that takes more than text as input or output: images, audio, video, PDFs. Gemini 3.1 Pro accepts all of them in one prompt, and Qwen3.8-Max accepts up to two hours of video. In chat apps multimodality shows up as file upload, voice mode and image generation; on pricing pages it shows up as separate per-minute or per-image rates.

On-device

Inference that happens entirely on your phone or computer, with nothing sent to a server. It is the strictest form of privacy and works offline, but only small models fit, so quality trails the cloud. Apple, Google and Meta all ship on-device models for tasks like transcription and summarizing notifications; Muse Glimmer's open weights are meant to run this way.

Open weights

The trained model file is published for anyone to download, inspect and run on their own hardware. Mistral (Apache 2.0), DeepSeek and Meta's Muse Glimmer do this; OpenAI, Anthropic and Google's flagships do not. It is not the same as open source: the training data and code usually stay private, and licenses vary. Open weights are why the same model can be bought from a dozen hosts at different prices.

Parameters

The numbers inside a model that training adjusts, loosely the model's size. Grok 4.7 is reported at 2.1 trillion; Mistral Medium 3.5 is 128 billion. More parameters generally means more capability and more cost to run, but mixture-of-experts designs and distillation have loosened the link, so parameter counts are a worse guide to quality than they were two years ago.

Prompt injection

An attack where text the model reads, rather than the person using it, changes its behavior. OWASP distinguishes direct injection, where a user's own prompt overrides the rules, from indirect injection, where a web page, email or file the model processes contains hidden instructions. It matters most for agents that browse and act; Google cited improved injection resistance as a headline feature of Gemini 3.8. Treat any content an agent reads as untrusted.

Quantization

Shrinking a model by storing its parameters with fewer bits of precision, so it needs less memory and runs faster with a small loss in quality. It is how a model that needs a data-center GPU at full precision becomes something a laptop can run. When you download a local model and see "4-bit" or "8-bit" in the file name, that is the quantization level.

RAG

Retrieval-augmented generation: instead of relying on what the model memorized in training, the system first searches your documents or the web, finds the relevant passages, and pastes them into the prompt so the model answers from them. It is how Notion AI answers questions about your workspace and how Perplexity cites sources. RAG reduces hallucination and lets a model use information newer than its training cutoff.

Rate limit

The cap on how much you can use a service in a given window, whether that is messages per day on a chat plan or requests per minute on an API. Consumer plans in 2026 are mostly sold as rate limits in disguise: "5x usage" means a five-times larger allowance in a rolling window, not a different product. The vendor's help page, not the pricing page, usually has the real numbers.

Reasoning model

A model trained to think before answering, spending extra tokens on hidden chain-of-thought and often checking its own work. ChatGPT's Think button, the effort slider on GPT-6 Astra and Claude's adaptive thinking are all reasoning controls. Reasoning models are slower and cost more per question, and they are the right tool for math, code, planning and any question with a checkable right answer.

Red teaming

Deliberately attacking a model before release to find what it will do wrong: extract private data, produce dangerous instructions, follow a prompt injection, or act outside its permissions. Labs run internal red teams and hire outside ones, and the results are supposed to appear in the model card. When OpenAI reported that Astra met its Critical cybersecurity threshold, that finding came from red teaming.

RLHF

Reinforcement learning from human feedback: after initial training, people rate pairs of model answers, and the model is trained to prefer the kind of answer people rated higher. It is the technique that turned raw next-word predictors into assistants that follow instructions and decline harmful requests. It also explains why models can sound agreeable rather than accurate; they were rewarded for answers people liked.

Synthetic data

Training data generated by a model rather than collected from people, used to fill gaps where real examples are scarce or private, such as rare medical cases or code in an unusual language. Labs increasingly train on it, and a model trained partly on another model's output is one reason distillation restrictions exist. The risk is that errors in the generating model get baked into the next one.

System prompt

The instructions the operator gives the model before your conversation starts: who it is, what it may and may not do, what tone to use. In chat apps it is hidden; in the API you write it yourself, and features like custom instructions and projects let you add to it. Most "the assistant will not do X" behaviors live here rather than in the model.

Temperature

An API setting that controls how random the model's word choices are. At zero it picks the most likely next token every time, giving consistent, repeatable answers; higher values make it more varied and, at the extreme, incoherent. Use low temperature for extraction and classification and higher for brainstorming. Chat apps hide it and pick a middle value.

Text-to-speech

Generating spoken audio from written text. Modern systems produce natural pacing and emotion rather than the flat voices of old screen readers, and they are priced per character or per minute; Mistral's Voxtral TTS is listed at €10 per million characters. ElevenLabs is the reference product, and Descript uses it to patch mistakes in a recording by typing.

Token

The unit a model reads and writes, roughly three-quarters of an English word, so "September" is one token and "unbelievably" may be three. Everything is priced and limited in tokens: input tokens are what you send, output tokens what the model writes, and a million of them is about 555,000 to 750,000 words depending on the tokenizer. A short email is around 200 tokens; a 300-page book around 100,000.

Tokens per minute

The most common API rate limit: how many tokens an account may send and receive each minute, usually rising with how much you have spent. Hitting it returns an error rather than a bill. It is separate from the context window, which caps a single request, and from monthly spend limits, which cap your money. xAI, OpenAI and Anthropic all publish tiered limits in their docs.

TPU

Tensor processing unit: Google's own chip for training and running models, an alternative to Nvidia GPUs. Google trains and serves Gemini on TPUs and rents them through its cloud, which is one reason it can price Flash models aggressively. For a user the chip is invisible; it matters when you read about supply constraints or a lab's compute deals.

Training data

The text, images, code and audio a model learned from, and the most contested subject in AI. It sets what the model knows, when its knowledge stops (the training cutoff, listed in model cards) and whose work it absorbed. On personal chat plans, your conversations become training data by default unless you opt out; on business plans and APIs they do not.

Transformer

The neural-network design, introduced in 2017, that underlies every current LLM and most image and audio models. Its key idea, attention, lets the model weigh every token in the context window against every other, which is what makes long-context understanding possible and also what makes long contexts expensive. When you see "GPT," the T is transformer.

Vector database

A database built to store embeddings and find the nearest ones quickly, so a RAG system can pull the ten most relevant paragraphs out of a million in milliseconds. Developers use products like Pinecone or the vector features inside Postgres; a chat app's "search your files" feature is a vector database with a friendly front end.

Voice cloning

Creating a synthetic voice that matches a specific real person from a short sample of their speech. It powers legitimate uses such as dubbing a creator's video into ten languages, and it is the tool behind voice-scam deepfakes. Reputable services require the speaker's consent and verification; ElevenLabs gates cloning behind both.

Watermarking

Embedding a signal in AI-generated content so it can later be identified as synthetic. Google's SynthID puts invisible watermarks into images, audio, text and video made by its models, and the C2PA standard attaches signed provenance data to files. Watermarks survive light editing and fail under heavy editing, so they are evidence, not proof, and only cover models whose makers chose to add them.

Zero-shot

Asking a model to do a task with no examples, just the instruction: "classify these reviews as positive or negative." Modern models are good enough at zero-shot that most everyday prompts are zero-shot without anyone noticing. When results are inconsistent, the fix is usually few-shot: add two examples and try again before reaching for anything more expensive.

Quick answers

What is a token in AI?

A token is the unit a language model reads and writes, roughly three-quarters of an English word. Prices and limits are quoted per token: a million tokens is about 555,000 to 750,000 words depending on the model, and a typical email is around 200 tokens.

What is the difference between a context window and memory?

The context window is how much text the model can hold in a single request, measured in tokens. Memory is a separate feature where the app stores facts about you between conversations and pastes them back into the context window later. A big context window does not give a model memory, and memory does not enlarge the window.

What does open weights mean?

The trained model file is published so anyone can download and run it, as Mistral, DeepSeek and Meta's Muse Glimmer do. It is not the same as open source, because the training data and code usually stay private, and the license may restrict commercial use.

More guides

• 
  Guides

  AI model tracker

  Every current model with its context window and price per token.

• 
  Guides

  How to choose an AI tool

  Five steps and a printable checklist, no month wasted.

• 
  Guides

  Best AI chatbot 2026

  Where these terms show up on the four big pricing pages.