Large Language Model. An AI model trained on vast text corpora to understand and generate natural language. Claude, ChatGPT, and Gemini are LLMs.

Unit of text processed by the model. ~1 token ≈ 0.55 word in English with the current tokenizer (0.75 before Opus 4.7). 1M tokens ≈ 555,000 words ≈ 7 novels.

The model's "working memory." Everything it can consider at once (your question + the history + the files).

The instruction you give Claude. A good prompt = a good result. It's the "brief" you'd give a consultant.

Standalone content created by Claude (app, dashboard, document) in a dedicated panel, editable and shareable.

An instructions file (SKILL.md) that specializes Claude in a specific competency. Reusable and shareable.

A complete package with skills + agents + connectors + configurations, installable in one command.

Model Context Protocol. Open standard for connecting Claude to external tools (Gmail, Slack, databases…).

Autonomous AI that carries out complex tasks using tools, APIs, and workflows, with minimal human intervention.

When the model generates false information stated with confidence. A structural risk of every LLM.

Anthropic's method: the model follows a built-in ethical "constitution" rather than simple filters.

Reinforcement Learning from Human Feedback. OpenAI's method: the model learns from the preferences of human raters.

Application Programming Interface. Technical interface for software to communicate with each other. The Claude API lets you embed AI in your tools.

A Claude Desktop mode that acts as an autonomous colleague: accesses your files, creates documents, carries out tasks.

A feature that lets you control Claude from your phone while it works on your computer. It is no longer offered to new users (existing users can keep it for now): you follow and start tasks from the Claude mobile app.

A mode where Claude "thinks" longer before answering, improving quality on complex tasks.

Anthropic's flagship model (June 2026), the first of the Mythos class open to the general public. State of the art on most benchmarks; excels at long, autonomous tasks. Replaced as the flagship by Fable 5.1 on September 1, 2026; still available.

Flagship model since September 1, 2026. Same price as Fable 5, but cache reads 4× cheaper (−25 to −45% on agentic workloads), finer-grained guardrails, and the first model to carry the text watermark. Since September 22, Opus 5.5 matches it on most tasks for far less.

A capability tier above the Opus class. Models so powerful they come with reinforced guardrails. Fable 5 and Mythos 5 are its first representatives.

The same model as Fable 5 but with certain guardrails lifted. Restricted access for partners of Project Glasswing (~150 organizations, 15+ countries: cyberdefense of critical infrastructure, biomedical research).

Model released July 24, 2026: close to Fable 5's intelligence for half the cost, with a distinctive habit of verifying its own work before delivering. Superseded by Opus 5.5 on September 22, 2026; still available (legacy).

Model released September 22, 2026, the first of the 5.5 family: Fable 5.1-level performance on most tasks for $4 / $20 per million tokens (40% less than Opus 5 on typical workloads), 30% faster. Thinking always on, five effort levels. Followed by Sonnet 5.5 (September 28) and Haiku 5.5 (October 7).

The everyday model since September 28, 2026: more than 30% faster than Sonnet 5 and up to 30% cheaper per task, at the same price ($2 / $10 per million tokens). Within 2 points of Opus 5.5 on GDPval-AA, a test of real professional tasks (per Anthropic). Open to everyone, Free included.

The smallest and fastest current model (October 7, 2026): about 75% cheaper than Haiku 4.5 on average, 1M context, the first Haiku with an adjustable effort level. Built for repetitive high-volume tasks: sorting, summaries, classification.

Fable 5's safety mechanism: on sensitive topics (cyber, bio/chem), the answer is automatically produced by an Opus model (4.8 or 5). Affects fewer than 5% of sessions, and even fewer with Fable 5.1. Since September 28, 2026, Sonnet 5.5 does the same on cybersecurity: it falls back to Sonnet 5 (can be turned off in Settings → Capabilities).

Invisible statistical mark in Claude's word choices (from Fable 5.1 on), verifiable by authorized bodies. Proves Claude was involved, not authorship. Cannot be disabled; generated files additionally carry C2PA metadata.

A standardized reference test used to compare AI models against each other (e.g. SWE-bench for code). Read with caution: no test sums up real-world use.

The "access badge" mechanism: you authorize Claude with the provider (Google, Slack…) without ever giving it your password. Revocable at any time.

Single Sign-On: your team members sign in to Claude with their corporate account (internal directory). SAML and SCIM are the technical standards that make it possible.

Data Processing Agreement: the data protection agreement required by Article 28 of the GDPR when a provider (here, Anthropic) processes data on your behalf.

A public trial version of a product still in development: functional, but liable to change quickly.