Guide

AI Glossary 2026: Every Term That Actually Matters, Explained

A plain-English AI glossary for 2026: more than 110 terms defined in one sentence each, then explained in context — from context window, mixture of experts and test-time compute to agentic AI, MCP, RAG, watermarking and the EU AI Act's compute thresholds. Every term links to the page that covers it in depth.

August 27, 2026 · The AI Rankings

Quick answer: This glossary defines the 110-plus AI terms that appear in real 2026 product pages, pricing tables and benchmark reports, one sentence each, in plain English. If you only learn ten, learn these: token, context window, parameter, inference, fine-tuning, RAG, hallucination, reasoning model, agent and open weights — between them they explain most of what a model page is telling you. The one caveat worth carrying into every vendor announcement: several of these words have no agreed technical definition and are used as marketing, particularly “agentic”, “open source”, “reasoning” and “AGI”, so this page says where the industry disagrees rather than pretending a consensus exists.

Terms are grouped A to Z. Each definition opens with a single self-contained sentence, then adds the context that makes it usable, and links to the page on this site that covers it properly. For the concepts that need a full explainer rather than a paragraph, see what is agentic AI and what is AI watermarking; for how the terms map onto actual products, see best AI models and best AI apps.

Start here: the ten terms that unlock the rest

If you read a model page and understand these ten, you can read any model page.

A

A2A (Agent2Agent)

A2A is an open protocol for agents built by different vendors to discover one another, exchange messages and coordinate tasks.

Google announced A2A on 9 April 2025 and donated it to the Linux Foundation on 23 June 2025; by its first anniversary in April 2026 more than 150 organisations backed it, and on 20 August 2026 it moved again, into the Agentic AI Foundation — the Linux Foundation body formed in December 2025 with MCP among its founding projects. Both protocols now sit under the same Linux Foundation-hosted body rather than with the companies that wrote them. See what is agentic AI for how the two fit together.

Agent

An AI agent is a program that wraps a language model in a loop and gives it tools, so it can take actions towards a goal rather than only produce a reply.

The loop is the defining feature: the agent decides on a step, calls a tool such as a web browser, a code interpreter or an API, reads the result, and decides what to do next. A chatbot answers a question and stops. An agent keeps going until the job is done or it gives up. See best AI agents for the products and best AI agent frameworks for the developer libraries used to build them.

Agentic AI

Agentic AI is the broader capability of pursuing a goal autonomously across multiple steps, and in its fuller form the orchestration of several agents into one system.

Treat “agentic” as a property with degrees, not a yes-or-no label. The useful question about any product is how much of the work it does unsupervised and where it stops to check with you. It is also the most inflated word in the 2026 vocabulary — Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027. Full explainer: what is agentic AI.

AGI (artificial general intelligence)

AGI is a hypothetical AI system that matches or exceeds human performance across essentially all cognitive tasks, rather than excelling in narrow domains.

There is no agreed technical definition and no agreed test, which is why lab announcements and sceptical rebuttals can both be internally consistent. When a company claims progress towards AGI, the only useful response is to ask which specific measurable capability improved. Compare superintelligence.

Alignment

Alignment is the field concerned with making an AI system pursue the goals its developers and users actually intend, including when those goals were never stated explicitly.

In practice it covers the post-training techniques that shape behaviour (RLHF, constitutional AI), the evaluations that test for unwanted behaviour, and the research into why models do what they do. It is distinct from safety engineering such as guardrails, which constrain a model from the outside.

API (application programming interface)

An API is the programmatic interface through which software sends a prompt to a model and receives a response, without any chat interface involved.

API access is how models get embedded in other products, and it is priced per token rather than per month. See best LLM APIs for the current provider comparison.

Artificial Analysis Intelligence Index

The Artificial Analysis Intelligence Index is an independent composite score that combines multiple public evaluations into a single number per model.

Version 4.1.1 combines nine evaluations — GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, AA-LCR, AA-Omniscience, Humanity’s Last Exam, GPQA Diamond and CritPt — weighted 34% agents, 24% coding, 24% scientific reasoning and 18% general (Artificial Analysis). It matters because it is run by a third party on a single harness, so scores are comparable across vendors in a way vendor-reported numbers are not. We use it as one input to best AI models, not as the ranking itself.

Attention

Attention is the mechanism that lets a model weigh how much each token in its input should influence each token it produces.

It is the core idea behind the transformer architecture, and the reason models can connect a pronoun to a noun forty paragraphs earlier. Its cost grows steeply with sequence length, which is why long context windows were an engineering achievement rather than a free parameter.

B

Base model

A base model is a model straight out of pre-training, before any instruction or safety training has been applied.

It predicts plausible continuations rather than answering questions, so it is rarely shipped to end users. Everything you interact with is a base model plus post-training.

Benchmark

A benchmark is a fixed set of tasks with known correct answers, used to score and compare models.

The two rules for reading them in 2026: check who ran it, and check whether the tasks have leaked into training data. Vendor-reported scores and independent scores on the same benchmark can differ by double digits — on SWE-bench Pro, vendor-reported figures run 17 to 21 points above the same models on Scale’s standardised harness. See SWE-bench, GPQA Diamond and Humanity’s Last Exam.

Bias

Bias in AI means systematic, unwanted skew in a model’s outputs, usually inherited from imbalances in its training data.

It is measurable and it is not the same as a model having opinions: a model that recommends men more often than women for the same job description is exhibiting bias whether or not it expresses any view about gender.

Blended price

Blended price is a single per-million-token figure that combines a model’s input and output prices at an assumed ratio, used to compare models on one axis.

Input and output tokens are almost never priced the same — Claude Opus 5 is $5 per million in and $25 per million out — so any single “price” quoted for a model is a blend, and the ratio assumed changes the answer.

C

C2PA

C2PA (Coalition for Content Provenance and Authenticity) is an open standard for attaching cryptographically signed metadata to a file recording how it was created and edited.

It is provenance labelling rather than watermarking: the signature lives in the file’s metadata, so it survives faithful copying but is removed by a screenshot or a metadata strip. OpenAI applies it to DALL-E 3 and Sora output, Anthropic applies it to generated .svg, .png and .jpg files, and Meta uses C2PA-style Content Credentials. Full treatment: what is AI watermarking.

Chain of thought

Chain of thought is the intermediate reasoning a model generates before its final answer, either because it was prompted to or because it was trained to.

Asking a model to “think step by step” was a prompting trick in 2022; by 2026 it is baked into reasoning models as trained behaviour. Note that the visible chain of thought is not a reliable account of the computation that produced the answer — it is text the model generated, and it can be wrong about its own process.

Chatbot

A chatbot is a conversational interface to a language model, which responds to each message and then waits.

The distinction from an agent is action: a chatbot produces text, an agent produces effects. Most 2026 consumer products are both, depending on which mode you invoke. See best AI chatbots.

Classifier

A classifier is a model that sorts inputs into categories rather than generating text, and in AI products it usually runs alongside the main model to flag or block content.

Frontier providers run classifiers as guardrails on both prompts and outputs. Their tuning is a real product difference: a tighter safety classifier can make an otherwise identical model refuse more often.

Compute

Compute is the quantity of processing work used to train or run a model, measured in FLOPs for training and in chip-hours for budgeting.

Training compute has become a regulatory unit as well as an engineering one — see compute threshold.

Compute threshold

A compute threshold is a regulatory line drawn at a specific quantity of training computation, above which a model attracts extra legal obligations.

Under the EU AI Act, a model trained with more than 10^23 FLOPs is presumed to be a general-purpose AI model, and one trained with more than 10^25 FLOPs is presumed to carry systemic risk, obliging the provider to notify the European Commission’s AI Office within two weeks (European Commission). Both presumptions are rebuttable: a provider can argue its model does not qualify.

Constitutional AI

Constitutional AI is a training method, developed by Anthropic, in which a model critiques and revises its own outputs against a written set of principles instead of relying solely on human preference labels.

It is one answer to the bottleneck in RLHF, which needs large volumes of human comparisons. The “constitution” is a published document, which makes the intended behaviour auditable in a way that a preference dataset is not.

Context rot

Context rot is the observed degradation in a model’s accuracy as the amount of material in its context window grows, even when the window is not full.

It matters because a headline context figure is a capacity, not a promise of uniform performance across that capacity. A model advertising a million tokens may still handle a 50,000-token document more reliably than a 500,000-token one.

Context window

The context window is the maximum number of tokens a model can consider at once, covering the system prompt, the conversation so far, any attached documents and the response being generated.

In August 2026 the frontier clusters at 1 million tokens — Claude Opus 5, GPT-5.6 Sol and Kimi K3 all sit there. The largest figures belong to models that are not frontier-available: Llama 4 is specified at 10 million tokens, and Gemini 3.5 Pro at 2 million but is still unreleased as of 27 August 2026. Bigger is not automatically better: see context rot.

Cost per completed task

Cost per completed task is the total spend required to get a job actually finished, rather than the headline price per million tokens.

It is the more honest comparison for reasoning models and agents, because a cheaper model that needs three attempts and more thinking tokens can cost more than an expensive model that succeeds first time. On best AI models the frontier’s lowest cost per completed task in August 2026 is Grok 4.6 at $0.84.

D

Data annotation

Data annotation is the human work of labelling, ranking, correcting or writing data that is then used to train or evaluate a model.

It is the labour layer underneath every frontier model, and it is a live job market rather than a historical footnote. See data annotation jobs and AI training jobs for what the work is and what it pays.

Dataset contamination

Dataset contamination is the leakage of benchmark questions and answers into a model’s training data, which inflates its score without improving its ability.

It is the reason benchmarks have a shelf life and why newer suites such as SWE-bench Pro deliberately include private codebases that cannot have been scraped. Treat any high score on an old, widely-published benchmark with caution.

Deepfake

A deepfake is synthetic audio, image or video that realistically depicts a real person doing or saying something they did not.

The EU AI Act treats deepfakes as a distinct category carrying an explicit disclosure duty on top of the general machine-readable marking requirement in Article 50.

Diffusion model

A diffusion model generates images, audio or video by starting from random noise and repeatedly removing noise until a coherent output emerges.

It is a different architecture from the transformer that powers text models, which is why image and text capabilities in one product often come from two separate models. See best AI image generators and best AI video generators.

Distillation

Distillation is training a smaller model to reproduce the outputs of a larger one, transferring much of the capability at a fraction of the running cost.

Most cheap, fast models in a family are distilled from the flagship — it is the standard reason a “mini” or “flash” tier can approach frontier quality on routine tasks. It is also a live legal question, since several providers’ terms prohibit using their outputs to train competing models.

E

Embedding

An embedding is a list of numbers representing a piece of text, image or audio, positioned so that similar meanings sit close together in that numerical space.

Embeddings are what make semantic search and RAG possible: you convert your documents and the user’s question into the same space and retrieve the nearest matches. They are stored and searched in a vector database.

EU AI Act

The EU AI Act is the European Union’s horizontal AI regulation, which applies obligations in tiers according to how risky a system’s use is.

The dates that matter as of 27 August 2026: obligations for providers of general-purpose AI models took effect on 2 August 2025; the Act’s general application and the Article 50 transparency duties began on 2 August 2026; and high-risk obligations were pushed back by the Digital Omnibus amendments to 2 December 2027 for Annex III systems and 2 August 2028 for product-embedded ones (artificialintelligenceact.eu). One deferral is easy to miss: for synthetic-content systems already on the EU market before 2 August 2026, the machine-readable marking obligation was postponed to 2 December 2026.

Eval (evaluation)

An eval is a test suite an organisation runs against a model to measure whether it does a specific job well enough, as distinct from a public benchmark.

The distinction matters commercially: public benchmarks tell you which models to shortlist, internal evals tell you which one works on your data. Tooling for this is a whole product category — see best LLM observability and eval tools.

Explainability

Explainability is the ability to give a human-understandable account of why a model produced a particular output.

It is narrower and more practical than interpretability, which asks what is happening inside the model. Note that a model’s stated reasons for an answer are generated text, not an audit trail, so an explanation is not automatically an explanation of the actual computation.

Export controls

Export controls are government restrictions on selling or transferring specified technology across borders, applied in AI to advanced chips and, increasingly, to model access itself.

They are no longer an abstraction for end users: Claude Fable 5 was suspended between 12 and 30 June 2026 under a US export-control directive before returning to general availability on 1 July, and its restricted twin Claude Mythos 5 remains trusted-access only.

F

Few-shot prompting

Few-shot prompting is including a handful of worked examples in the prompt so the model infers the pattern you want.

It remains the cheapest way to control output format, and it is often more effective than a longer instruction. Compare zero-shot.

Fine-tuning

Fine-tuning is additional training of an already-trained model on a narrower dataset, to specialise its behaviour, tone or domain knowledge.

It changes the model’s weights, which distinguishes it from RAG, which changes what the model is shown. The practical rule: fine-tune to change behaviour and style, use RAG to supply facts. See also LoRA.

FLOP

A FLOP is one floating-point operation, and FLOPs are the standard unit for measuring how much computation a training run consumed.

The numbers are large enough that only the exponent matters in practice — 10^25 FLOPs is a hundred times 10^23. See compute threshold.

Foundation model

A foundation model is a large model trained on broad data that serves as the base for many downstream applications rather than one specific task.

The term is largely interchangeable with “base model” and, for text, with large language model. Regulators prefer “general-purpose AI model”, which is the term used in the EU AI Act.

Frontier model

A frontier model is one of the small number of models at the leading edge of general capability at a given moment.

On this site, Frontier is a defined tier rather than a loose compliment: as of August 2026 it covers Claude Mythos 5, Claude Fable 5 and Claude Opus 5, with Flagship, Strong, Value and Open tiers beneath. The full tiering and its evidence sit on best AI models.

G

Frontier safety framework

A frontier safety framework is a provider’s published policy committing it to specific capability evaluations and to defined safeguards once a model crosses a stated risk threshold.

Each major lab maintains its own version under its own name, and the thresholds are self-set rather than regulated, so the documents are commitments rather than compliance. They are still the clearest statement of what a provider believes its models can do.

Function calling

Function calling is the mechanism by which a model returns a structured request to run a named function with specific arguments, instead of returning prose.

It is the low-level plumbing beneath tool use: the model does not execute anything itself, it emits a request that the surrounding application executes and feeds back. MCP standardises how those functions are described and discovered.

Generative AI

Generative AI is AI that produces new content — text, images, code, audio or video — in response to a prompt.

It is the reactive layer beneath everything else: it does nothing until asked, and it stops once it has answered. An agent is generative AI plus a loop plus tools.

GPQA Diamond

GPQA Diamond is a set of 198 graduate-level science questions written to be difficult for non-experts to answer even with unrestricted web access.

It was the standard hard-reasoning test until it saturated: by 2026 the top models cluster at 93 to 95%, so it separates almost nothing at the frontier and has largely been superseded by Humanity’s Last Exam and harder agentic evals.

GPU

A GPU (graphics processing unit) is the parallel processor that does the arithmetic in AI training and inference.

GPU supply, not algorithms, is usually the binding constraint on how fast a lab can train and how cheaply it can serve. The main alternative is Google’s TPU.

Grounding

Grounding is tying a model’s output to a verifiable external source, such as a retrieved document or a live search result, rather than to its internal parameters alone.

It is the mechanism behind citations in AI search engines, and it reduces hallucination without eliminating it, because a model can still misread or misattribute a source it was given.

Guardrails

Guardrails are external checks that sit around a model, filtering inputs or outputs against rules the model itself is not trusted to enforce.

They are engineering controls, not alignment: a guardrail blocks a bad output after the fact, whereas alignment aims at a model that does not produce it. Most production systems use both.

H

Hallucination

A hallucination is a model output that is fluent, confident and factually wrong.

The mechanism is not a bug in the ordinary sense: a language model is optimising for plausible continuations, and a plausible-sounding citation is exactly what it is built to produce. This is why grounding, RAG and verification exist, and why hallucination rates are now a scored component of composite indices — AA-Omniscience contributes a specific non-hallucination score to the Artificial Analysis Intelligence Index.

Human-in-the-loop

Human-in-the-loop describes a system design in which a person reviews or approves the AI’s work at defined points rather than after the fact.

It is the honest description of most “autonomous” 2026 products, and the checkpoints are where the real design work lies.

Humanity’s Last Exam

Humanity’s Last Exam is a benchmark of expert-level questions across many disciplines, built deliberately to be hard enough to remain unsaturated as models improve.

It carries 12% of the Artificial Analysis Intelligence Index and has become the headline “hard reasoning” number now that GPQA Diamond has saturated.

I

Inference

Inference is running a trained model to produce an output, as opposed to training it.

Every consumer subscription and every API bill is paying for inference. It is priced per token, split between input and output, and it is where the cost of a deployed AI system actually lives.

Instruction tuning

Instruction tuning is post-training a base model on examples of instructions and good responses, so it follows requests rather than merely continuing text.

It is the step that turns a raw foundation model, which will happily continue your question as if it were a document, into something that answers it.

Interpretability

Interpretability is the research field trying to explain what is happening inside a model in terms humans can check.

It matters commercially as well as scientifically: an uninterpretable model is one whose failures you can only discover empirically, which is a poor fit for regulated deployments.

J

Jailbreak

A jailbreak is a prompt crafted to make a model bypass its own safety training and produce output it was trained to refuse.

It is distinct from prompt injection: a jailbreak is the user attacking the model’s rules, while a prompt injection is a third party attacking the user through content the model reads.

K

Knowledge cutoff

A knowledge cutoff is the date after which a model’s training data ends, and therefore the point past which it has no built-in knowledge of events.

It does not limit what the model can discuss — grounding and web search supply anything later — but it does determine what it will get wrong when it answers from memory. A model asked about a product released after its cutoff will often answer confidently from an older version.

L

Latency

Latency is the delay between sending a request and receiving output, usually split into time-to-first-token and generation speed.

It is the main practical cost of reasoning models, which can spend tens of seconds on test-time compute before the first word appears.

LLM (large language model)

A large language model is a model trained on very large volumes of text to predict the next token, which turns out to produce general language capability.

Almost every product on this site is an LLM or a wrapper around one. The category has widened — most 2026 frontier LLMs are multimodal — but the name has stuck.

LMArena

LMArena is a public evaluation platform where people compare two anonymous model responses side by side and vote for the better one, producing a preference-based ranking.

It measures what humans prefer, which correlates with but is not the same as what is correct. It is a useful counterweight to benchmark scores and a poor substitute for them.

LoRA (low-rank adaptation)

LoRA is a fine-tuning technique that trains a small set of additional parameters and leaves the original model’s weights frozen.

It cuts the cost and hardware requirement of fine-tuning by orders of magnitude, and it is why running many task-specific variants of one base model is practical.

M

MCP (Model Context Protocol)

MCP is an open standard for connecting AI models to external tools and data sources through one common interface, so a given integration works across any client that speaks the protocol.

Anthropic introduced it and then donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI and supported by Google, Microsoft, AWS, Cloudflare and Bloomberg (Anthropic). Its legal home is now “Model Context Protocol a Series of LF Projects, LLC”, with contributions under Apache 2.0. In one sentence: MCP standardises how one agent reaches its tools, while A2A standardises how agents reach each other.

Memory

Memory is a product feature that lets an assistant retain information across separate conversations, as opposed to the within-conversation recall provided by the context window.

It is implemented as storage plus retrieval, not as a change to the model, which is why memory can be edited or switched off. Availability varies by plan — see best AI apps.

Mixture of experts (MoE)

Mixture of experts is an architecture that splits a model into many sub-networks and activates only a few of them per token, so total parameter count and per-token compute cost are decoupled.

This is why 2026 parameter counts stopped being a useful capability signal: a sparse model with a trillion total parameters may activate a small fraction of them on any given token, running far cheaper than a dense model of the same nominal size.

Model card

A model card is a short structured document describing a model’s intended uses, limitations, training data at a high level and evaluation results.

It is the lighter-weight sibling of a system card and, under the EU AI Act, overlaps with the technical documentation general-purpose model providers must maintain.

Model collapse

Model collapse is the degradation that occurs when models are trained repeatedly on model-generated output, causing the distribution of the data to narrow with each generation.

It is the technical reason human-produced data annotation has not been automated away, and it is why provenance signals such as watermarking matter to labs as well as to regulators.

Multi-agent system

A multi-agent system is an arrangement in which several specialised agents coordinate on one objective, usually with an orchestrator dividing the work.

It is the shape most enterprise “agentic” deployments actually take, and the reason A2A exists as a standard.

Multimodal

A multimodal model accepts or produces more than one type of data — typically some combination of text, images, audio and video.

By 2026 multimodal input is standard at the frontier and multimodal output is not: most flagship text models read images fluently but generate images through a separate diffusion model, if at all.

N

Natural language processing (NLP)

Natural language processing is the broader computer-science field concerned with getting machines to work with human language, of which today’s language models are one approach.

The term predates the current era by decades and still covers tasks — parsing, entity extraction, sentiment classification — that are often better served by small specialised models than by a frontier LLM.

Neural network

A neural network is a model made of layers of simple numerical units whose connection strengths are learned from data.

Every term on this page that involves training, parameters or weights is describing some property of a neural network. The transformer is the particular design that made language models work.

Non-determinism

Non-determinism is the property that the same prompt can produce different outputs on different runs, even at a fixed temperature.

It comes from sampling and from hardware-level floating-point variation, and it is why AI features need evals over many runs rather than a single passing test.

O

Open source AI

Open source AI, used strictly, means a model whose weights, training code and training data are all released under a licence permitting use, study, modification and redistribution.

Very few widely-cited “open source” models meet that bar, because training data is almost never released. The Open Source Initiative’s definition and common industry usage diverge sharply here, which is why this site prefers the narrower, checkable term open weights. See best open-source LLMs.

Open weights

Open weights means the trained model file is downloadable and runnable on your own hardware, whatever the licence attached to it.

It is a factual claim about availability, not a licensing claim: some open-weight models carry genuinely permissive licences (DeepSeek V4 is MIT), others carry use restrictions. Open weights is what determines whether you can self-host — see best local LLMs.

Orchestration

Orchestration is the layer that decides which model, tool or agent handles each part of a task and in what order.

It is where most of the engineering effort in a production AI system now sits, and it is what an agent framework provides — see best AI agent frameworks.

Output token

An output token is a token the model generates, as opposed to one it reads, and it is billed at a higher rate than an input token on essentially every provider.

The multiple is typically five to six times, with a range of roughly two to six across providers: Claude Opus 5 is $5 per million input tokens and $25 per million output, and Gemini 3.1 Pro is $2 and $12. Reasoning models bill their thinking as output tokens, which is why a single hard question can cost more than a long document upload.

P

Parameter

A parameter is one learned numerical value inside a model, and parameter count is the traditional shorthand for model size.

It has become a weak capability signal, for two reasons: mixture of experts decouples total parameters from active compute, and post-training now accounts for a large share of a model’s usefulness. Most frontier labs stopped publishing parameter counts.

Pass@1

Pass@1 is a scoring rule that counts a task as solved only if the model gets it right on its first attempt.

It is the standard used by SWE-bench Pro and by most of the evaluations inside the Artificial Analysis Intelligence Index, and it is a much harder bar than pass@k, which allows several attempts. Always check which one a quoted score used.

Post-training

Post-training is everything done to a model after the main pre-training run — instruction tuning, RLHF, safety training and reasoning training.

The gap between a base model and a shipped product is mostly post-training, which is why two models built on comparable pre-training can behave very differently.

Pre-training

Pre-training is the initial, expensive training run in which a model learns language and world knowledge from a very large corpus by predicting the next token.

It is the phase that consumes the FLOPs counted by regulatory compute thresholds.

Prompt

A prompt is the full input given to a model for one turn, including the user’s message, the system instructions and any attached content.

Everything in the prompt occupies the context window and is billed as input tokens.

Prompt engineering

Prompt engineering is the practice of designing prompts and prompt structures to get reliable output from a model.

It survived its own hype cycle: as a standalone job title it has largely been absorbed into AI engineering roles, though the skill remains valuable. For what the market actually pays, see prompt engineer salary.

Prompt injection

Prompt injection is an attack in which instructions hidden inside content the model reads — a web page, an email, a document — hijack its behaviour.

It is the central unsolved security problem for agents, because an agent that browses the web is by definition reading untrusted instructions. Unlike a jailbreak, the victim is the user, not the model’s policy.

Q

Quantisation

Quantisation is reducing the numerical precision of a model’s weights, shrinking memory requirements and speeding up inference at some cost to accuracy.

It is what makes running a large open-weight model on consumer hardware feasible: a model that needs 140GB at full precision may fit in 40GB when quantised. See best local LLMs.

R

RAG (retrieval-augmented generation)

RAG is the pattern of retrieving relevant documents from an external store and inserting them into the prompt so the model answers from them rather than from memory.

It is the standard architecture for making a general model useful on private, proprietary or recent data, and it is built from embeddings plus a vector database plus the model. It reduces hallucination but does not remove it: a model can still misread a correctly retrieved document.

Rate limit

A rate limit is a cap on how many requests or tokens an account may use in a given window, applied by every provider on both consumer plans and APIs.

It is the real constraint on heavy usage far more often than model capability is, and it is the substantive difference between most consumer subscription tiers.

Reasoning model

A reasoning model is one trained to generate extended intermediate steps before answering, spending more computation at inference time in exchange for higher accuracy on hard problems.

The trade is explicit: better results on maths, code and multi-step analysis, at higher latency and higher cost per answer. Most 2026 frontier models expose this as an adjustable effort setting rather than a separate model. See test-time compute.

Red teaming

Red teaming is deliberately adversarial testing, in which people or automated systems try to make a model behave badly before it ships.

Findings usually appear in the system card, and under the EU AI Act adversarial testing is part of the expected practice for models presumed to carry systemic risk.

Reward model

A reward model is a separate model trained on human preference data to score outputs, used to guide the main model during RLHF.

Its quality caps the quality of the alignment: a reward model that prefers confident-sounding answers will train a model that produces them, correct or not.

RLHF (reinforcement learning from human feedback)

RLHF is a post-training method in which humans rank alternative model outputs, a reward model learns those preferences, and the model is optimised against that reward.

It is the technique that turned capable-but-unruly base models into usable assistants, and it is the reason data annotation is a large employment category. Its main constraint is the volume of human judgement required, which is what constitutional AI and AI-assisted feedback methods try to relieve.

Router

A router is a component that decides which model handles a given request, typically sending easy queries to a cheap model and hard ones to an expensive one.

Consumer products increasingly do this invisibly, which is why the model you think you are talking to may not be the model that answered.

S

Scaffolding

Scaffolding is the code around a model in an evaluation or product — the retries, tool access, planning steps and prompts — that substantially changes measured performance.

It is the single biggest reason benchmark scores disagree. The same model, run on a vendor’s scaffold and on a standardised harness, can differ by 17 to 21 points on SWE-bench Pro. When comparing two scores, confirm they came from the same harness before drawing a conclusion.

Self-hosting

Self-hosting is running a model on infrastructure you control rather than calling a provider’s API.

It requires open weights, enough hardware to hold the model — often after quantisation — and someone to operate it. The usual reasons are data residency, regulatory constraint and unit cost at high volume. See best local LLMs.

Semantic search

Semantic search is retrieval by meaning rather than by keyword match, achieved by comparing embeddings of the query and the documents.

It is the retrieval half of RAG, and it is why a search for “how do I cancel” can return a document that only ever says “terminating your subscription”.

Structured output

Structured output is a mode in which the model is constrained to return data matching a specified schema, typically JSON, rather than free text.

It is what makes models usable as components in software rather than as chat interfaces, because downstream code can parse the result without guessing.

Superintelligence

Superintelligence is a hypothetical system substantially exceeding the best human performance across essentially all cognitive domains.

Like AGI it has no operational definition, and claims about its timeline are forecasts rather than measurements.

SWE-bench

SWE-bench is a family of benchmarks that scores a model on resolving real software issues from real repositories, verified by whether the repository’s own tests pass.

Two variants matter in 2026. SWE-bench Verified is the older, human-validated Python-only set, now partially contaminated and largely saturated. SWE-bench Pro, from Scale AI, contains 1,865 tasks across 41 repositories in Python, Go, TypeScript and JavaScript, split into public, commercial and held-out sets, with private codebases that models cannot have trained on (Scale SEAL). Always check which variant a score refers to and who ran it. See best AI for coding.

Synthetic data

Synthetic data is training or evaluation data generated by a model rather than collected from the world.

It solves data scarcity and privacy problems and introduces a new failure mode: training repeatedly on model-generated output degrades quality, the phenomenon usually called model collapse. See also best AI synthetic research for the market-research application of the same idea.

SynthID

SynthID is Google DeepMind’s watermarking family, covering text, images, audio and video, and now the de facto industry scheme for text.

SynthID-Text was published in Nature and open-sourced in October 2024, and on 15 August 2026 Anthropic confirmed Claude’s text watermark uses the same approach. Full explainer: what is AI watermarking.

System card

A system card is a detailed document published alongside a major model release covering capabilities, evaluations, safety testing and known limitations.

It is the most useful primary source about a new model, and it is worth reading against the marketing: Anthropic’s own System Card for Claude Opus 5 describes it as “not more capable overall” than the Mythos-class models above it.

System prompt

A system prompt is the standing instruction set given to a model before the user’s message, defining its role, constraints and available tools.

It occupies context and is billed like any other input, which is why long system prompts have a real per-request cost.

T

Temperature

Temperature is a setting that controls how much randomness a model uses when choosing among likely next tokens.

Low temperature gives repeatable, conservative output suited to extraction and code; higher temperature gives more varied output suited to brainstorming. It does not control accuracy, and turning it to zero does not prevent hallucination.

Test-time compute

Test-time compute is computation spent while answering rather than while training, and scaling it is the main capability lever of the 2025 to 2026 period.

It is the mechanism behind reasoning models: a model given more tokens to think produces better answers on hard problems, with the cost showing up on your bill per query instead of in a one-off training run.

Throughput

Throughput is how many tokens a model generates per second once it has started responding.

Together with time to first token it determines perceived speed, and the two can pull in opposite directions: a reasoning model may generate quickly but start slowly.

Time to first token

Time to first token is the delay between sending a request and the first piece of the response appearing.

It is the number users actually feel, and it is where reasoning models pay for their accuracy — thinking happens before the first visible word.

Token

A token is the unit of text a model actually processes, typically a word fragment averaging about four characters or three-quarters of an English word.

Everything is denominated in tokens: context windows, API pricing, rate limits and throughput. As a rough conversion, 1,000 tokens is about 750 words, so a 100,000-word book is roughly 133,000 tokens.

Tokenisation

Tokenisation is the process of splitting text into tokens according to a fixed vocabulary learned during training.

It explains several odd model behaviours: poor arithmetic on long numbers, difficulty counting letters in a word, and the fact that non-English text often costs more per equivalent word because it fragments into more tokens.

Tool use

Tool use is a model’s ability to call external functions — search, code execution, database queries, third-party APIs — and incorporate the results into its answer.

It is the capability that separates an agent from a chatbot, and MCP is the standard for wiring it up.

TPU

A TPU (tensor processing unit) is Google’s custom AI accelerator chip, used as an alternative to GPUs for training and inference.

Its relevance to buyers is indirect but real: control of its own silicon is part of why Google can price Gemini the way it does.

Training data

Training data is the corpus a model learns from, and its composition determines the model’s knowledge, its biases and much of its legal exposure.

Under the EU AI Act, providers of general-purpose models must publish a sufficiently detailed summary of the content used for training. Full datasets are essentially never released, which is why almost no model meets the strict definition of open source AI.

Training run

A training run is one complete execution of a model’s pre-training, typically consuming months of chip time and tens of millions of dollars.

Its size in FLOPs is what regulatory compute thresholds measure.

Transformer

The transformer is the neural network architecture, introduced in the 2017 paper “Attention Is All You Need”, that underpins essentially every large language model in use today.

Its innovation was attention, which let models process a whole sequence in parallel while still weighing long-range relationships. Every model named on best AI models is a transformer variant.

V

Vector database

A vector database is a data store built to hold embeddings and retrieve the nearest matches to a query vector quickly.

It is the retrieval engine underneath most RAG systems, and it is a competitive product category in its own right.

Vibe coding

Vibe coding is building software by describing what you want in natural language and accepting the model’s code largely without reading it.

The term was coined by Andrej Karpathy in February 2025 and has since acquired a second, less flattering meaning: code nobody on the team can explain. It is a legitimate technique for prototypes and a documented source of production incidents. See best AI for coding and best AI code review tools.

Vision-language model (VLM)

A vision-language model is one that takes both images and text as input and reasons across them.

It is what lets you photograph a whiteboard, a spreadsheet or an error screen and ask a question about it. In 2026 this is standard at the frontier rather than a differentiator.

W

Watermarking

Watermarking is embedding a statistically detectable signal into a model’s output so it can later be identified as machine-generated.

For text it works by biasing the model’s choice among near-equivalent words using a secret key, which means it needs length to be detectable and degrades as text is rewritten. As of August 2026 Anthropic marks Claude’s text output worldwide and Google marks Gemini’s; OpenAI built a text watermarker and chose not to ship it. Full explainer: what is AI watermarking, and for the detection side, best AI detectors and best AI humanizers.

Weights

Weights are the learned numerical values that constitute a trained model, and the model file is essentially a very large table of them.

“Releasing the weights” means publishing that file — see open weights.

Z

Zero-shot

Zero-shot means asking a model to perform a task with no worked examples in the prompt, relying entirely on what it learned in training.

Frontier models are strong zero-shot, which is why prompting has become less finicky than it was. Compare few-shot prompting.

Related reading

For the concepts that need more than a paragraph, see what is agentic AI and what is AI watermarking. For how these terms attach to real products, see best AI models, best AI apps and best AI chatbots. For the developer stack, see best LLM APIs, best open-source LLMs, best local LLMs and best LLM observability and eval tools. For the human labour underneath the models, see AI jobs and data annotation jobs.

Frequently asked questions

What is an AI glossary?

An AI glossary is a reference list of artificial intelligence terms with short definitions. This one covers more than 110 terms in use in 2026, grouped A to Z, with a one-sentence definition first and the practical context second. It is written for someone who already uses AI tools and wants to read a model page, a pricing table or a benchmark report without guessing at the vocabulary.

What AI terms should a beginner learn first?

Ten terms cover most of what a beginner needs: token, context window, parameter, inference, fine-tuning, RAG, hallucination, reasoning model, agent and open weights. Between them they explain how a model is sized, how it is priced, how it is adapted to a specific job, how it fails, and what you can and cannot run yourself. Everything else in this glossary is a refinement of one of those ten ideas.

What is the difference between a token and a word?

A token is the unit of text a model actually processes, and it is usually smaller than a word — roughly four characters, or about three-quarters of an English word on average. Common words are single tokens, while rarer words, long numbers and non-English text split into several. The practical conversion is that 1,000 tokens is approximately 750 words, so a 100,000-word manuscript is roughly 133,000 tokens of context.

What is the difference between an AI agent and a chatbot?

A chatbot produces a reply and stops; an agent runs a loop, uses tools and keeps working until the goal is met. The technical difference is tool use and iteration: an agent can search the web, run code, call an API or edit a file, read what happened, and decide its next step. In 2026 most consumer AI products are both, switching into agent behaviour when you give them a task rather than a question.

What is the difference between fine-tuning and RAG?

Fine-tuning changes the model’s weights by training it further on your data; RAG leaves the model unchanged and instead retrieves relevant documents and puts them in the prompt. The practical rule is to fine-tune when you want to change how a model behaves — its tone, its output format, its domain style — and to use RAG when you want it to know facts it was never trained on. Many production systems use both, and RAG is the cheaper starting point because it requires no training run.

Does open source AI mean the same thing as open weights?

No. Open weights means the trained model file is downloadable and you can run it on your own hardware. Open source, used strictly, additionally requires the training code and training data to be released under a licence allowing use, modification and redistribution. Almost no widely-cited “open source” model meets the strict definition, because training data is essentially never published, so open weights is the more accurate and more checkable term.

Why do the same AI benchmark scores differ between sources?

Because the scaffolding differs. The code around the model in an evaluation — retries, tool access, planning steps, prompt structure — materially changes measured performance, so a vendor running its own harness and an independent lab running a standardised one will report different numbers for the same model. On SWE-bench Pro the gap runs 17 to 21 points in the vendor’s favour. Both numbers can be honest; treat vendor scores as a ceiling and standardised leaderboards as a floor, and never compare across harnesses.

What is a context window and how big are they in 2026?

A context window is the maximum number of tokens a model can hold in view at once, covering the system prompt, the conversation, any attached files and the answer being written. As of 27 August 2026 the frontier clusters at 1 million tokens, and the larger published figures belong to models outside that cluster: Llama 4 is specified at 10 million tokens, and Gemini 3.5 Pro at 2 million but is still unreleased. A larger window is not automatically better, because accuracy tends to degrade as more material is loaded in, a phenomenon known as context rot.

What does the EU AI Act require as of 2026?

As of 27 August 2026, the EU AI Act’s general application and its Article 50 transparency duties are in force, having taken effect on 2 August 2026, and obligations for general-purpose AI model providers have applied since 2 August 2025. High-risk system obligations were deferred by the Digital Omnibus amendments to 2 December 2027 for Annex III systems and 2 August 2028 for product-embedded ones. For synthetic-content systems already on the EU market before 2 August 2026, the machine-readable marking requirement was postponed to 2 December 2026.

What is MCP and how is it different from A2A?

MCP, the Model Context Protocol, standardises how one AI agent connects to external tools and data sources, so an integration written once works with any client that speaks the protocol. A2A, Agent2Agent, standardises how separate agents discover one another and coordinate work across vendors. In short, MCP connects an agent to its tools and A2A connects agents to each other. As of August 2026 both are governed by the Agentic AI Foundation, a Linux Foundation body, rather than by Anthropic and Google respectively.

Is AGI a technical term?

Not in any operational sense. Artificial general intelligence describes a system matching or exceeding human performance across essentially all cognitive tasks, but there is no agreed definition and no agreed test, which is why credible people disagree about how close it is without contradicting each other on any measurable fact. When a company claims progress towards AGI, the only informative question is which specific capability improved and on what evaluation.

What is prompt injection and why is it a problem for AI agents?

Prompt injection is an attack in which instructions hidden inside content a model reads — a web page, an email, a shared document — take over its behaviour. It is the central unsolved security problem for agents, because an agent that browses the web or reads a mailbox is by definition consuming untrusted instructions while holding access to your accounts and files. It differs from a jailbreak, where the user is the one trying to bypass the model’s rules; with prompt injection the user is the victim.

← All guides