What Is Claude AI? The Complete Guide

The map — everything in this guide

Expand all Collapse Fit Fullscreen − +

• Claude
  • What it is
    • An assistant built on the Claude models
    • Made by Anthropic
    • First released March 2023
    • Developer- and enterprise-weighted
  • The models
    • Haiku — fastest
    • Sonnet — the balanced default
    • Opus — agentic coding
    • Fable — the frontier
    • Adaptive thinking
    • 1m-token context
  • What it costs
    • Free
    • Pro — $17–20/mo
    • Max — 5× and 20×
    • Team & Enterprise
    • API per token
  • What makes it different
    • Artifacts
    • Claude Code
    • MCP
    • Projects
    • Constitutional AI
    • Willing to disagree with you
  • Getting good output
    • Instruction before data
    • Show one example of good
    • The prompt library
    • Advanced moves
  • Limits
    • Hallucination
    • Knowledge cutoff
    • Usage caps bite
    • Smaller consumer ecosystem
  • Claude vs the others
    • vs ChatGPT
    • vs Gemini
    • Test on your own work
  • Privacy
    • Consumer opt-out
    • Commercial excluded by default
  • Why it matters for your site
    • Claude cites sources
    • GEO
  • Learn it properly
    • Eight AI Hub tracks
    • Free, no sign-up

Click a node to open or close it · drag to pan · ↗ jumps to the section

What is Claude?

Claude is an AI assistant made by Anthropic — a chat interface, a set of desktop and mobile apps, a coding agent and an API, all built on a family of large language models also called Claude. You give it a task in plain language; it reads everything in front of it and writes back, one small piece of text at a time.

That description fits every assistant on the market, so here is the part that does not. Claude was built by a company founded specifically around the question of how to make powerful AI systems behave well, and that priority shows up in the product in ways you can feel within a few conversations. Claude is more willing to tell you that you are wrong. It is more careful about claims it cannot support. It is unusually strong at long documents, sustained reasoning and code. And it has spent its whole life over-indexed on developers and enterprises rather than on mass-market consumer reach.

If ChatGPT is the assistant that reached everyone, Claude is the one that a disproportionate number of people who build things reached for. By 2026 that split is visible in the numbers: Claude's web traffic is far more concentrated in the United States and in developer-heavy markets, while its revenue is weighted toward API and enterprise use rather than consumer subscriptions.

Two distinctions to fix before we go further:

• Claude is the product and the model family. Unusually, the name covers both. "Claude Opus 5" is a model; "Claude" in daily speech is the app you talk to.
• Claude is not a search engine or a database. Like every large language model, it generates text by prediction. It can search the web as a tool and cite what it finds, but it does not look facts up by default, and it can be confidently wrong.

For the machinery underneath — tokens, transformers, training, why models hallucinate — read our companion guide to large language models. This guide is about Claude as a working tool.

Who makes Claude, and why it matters

Anthropic was founded in 2021 by a group that included several senior people from OpenAI, among them the siblings Dario and Daniela Amodei. It is structured as a public-benefit corporation, which means its charter obliges it to weigh a public mission alongside shareholder returns — a legal detail that turns out to explain a lot of product decisions.

The founding thesis was that frontier AI was going to be built regardless, and that a safety-focused lab should be at the frontier rather than commenting from the sidelines. Two artefacts of that thesis matter to anyone using the product:

Constitutional AI. Rather than shaping a model's behaviour purely through human raters choosing between responses, Anthropic gives the model a written set of principles — a "constitution" — and trains it to critique and revise its own outputs against them. The practical effect is a model whose refusals and hedges follow a stated, published rationale rather than an opaque accumulation of preferences.

The Responsible Scaling Policy. A published framework of capability thresholds and the safeguards that must be in place before a model crossing one is deployed. You can disagree with where the lines sit; the useful part is that they are written down, which makes them arguable.

None of this makes Claude safer to use uncritically. It does mean that when Claude declines something, or flags uncertainty, or pushes back on your framing, it is doing something the company deliberately built rather than something that leaked out of the training data.

A short history of Claude

2021. Anthropic founded. No public product for two years; the first work is research, including the Constitutional AI method.

March 2023. Claude 1 ships, initially to a waitlist. It is immediately noted for long-context handling and a distinctly careful writing voice.

2023–2024. The Claude 2 and Claude 3 generations arrive, and Anthropic settles on the naming that has survived since: three sizes named after poetic forms — Haiku (fastest), Sonnet (balanced), Opus (most capable). Claude 3 is the point at which Claude becomes a serious daily alternative rather than an interesting second opinion.

2024. Artifacts ship — a side panel where Claude's output becomes a live document, app or diagram you can use rather than a block of text you copy. It is the first genuinely novel interface idea in assistant products since the chat box itself.

Late 2024. MCP, the Model Context Protocol, is published as an open standard for connecting assistants to tools and data. It is adopted well beyond Anthropic and becomes the default way assistants talk to external systems.

2025. Claude Code arrives: an agent that runs in your terminal, reads your repository, edits files, runs tests and iterates. It moves Claude from "helps you write code" to "does the work and shows you the diff."

2026. The Claude 5 generation — Sonnet 5, Opus 5 and the frontier Fable 5 — brings a one-million-token context window across the line and adaptive thinking that is always on at the top. A separate model, Mythos 5, ships invitation-only for defensive cybersecurity work. Alongside them, a set of task-shaped surfaces: Claude Code for engineering, Cowork for general knowledge work, Design, and Science.

The Claude model lineup in 2026

Four models are generally available, named on a scale from fastest to most capable. Unlike most competitors, Anthropic publishes the full specification of each — context, output ceiling, price, knowledge cutoff — which makes choosing between them unusually straightforward.

Two footnotes worth knowing. Sonnet 5 has carried introductory pricing of $2 / $10 per million tokens into late 2026, which makes it one of the best capability-per-dollar options on the market while it lasts. And a fifth model, Claude Mythos 5, shares Fable's specification but is offered invitation-only for defensive cybersecurity work under a programme called Project Glasswing — it is not something you can sign up for.

Adaptive thinking and effort

Older assistants shipped a separate "reasoning model" you had to select. Current Claude models instead use adaptive thinking: the model decides how much internal deliberation a request deserves and spends it. On Fable 5 it is always on; on Opus 5 and Sonnet 5 it is available and enabled by default in most surfaces.

Layered on top is an effort control — low, medium or high — which caps how much of that deliberation the model may spend. High effort is the default on the API and in Claude Code. Lower effort is faster and cheaper and perfectly adequate for straightforward work.

The practical translation for everyday use: you rarely need to manage this, and when you do, it is because you want it turned down, not up. If Claude is over-thinking a simple rewrite, say so. If it is rushing a genuinely hard problem, tell it the answer needs to satisfy specific constraints and it will spend more.

The million-token context window

Sonnet 5, Opus 5 and Fable 5 all hold roughly one million tokens at once — on the order of 555,000 words, or several full-length books. This is not a specification-sheet flourish; it changes what kinds of task are possible.

• Whole-repository reasoning. A mid-sized codebase fits. Claude can answer "where is this behaviour actually implemented" instead of guessing from a snippet.
• Whole-document review. A contract, a research corpus, a year of meeting notes, a full technical specification — analysed together rather than chunked and stitched.
• Long, coherent sessions. A working session can run for hours without the beginning falling out of memory.

Two cautions. First, a large window is not free: everything in it is billed and everything in it competes for the model's attention. Dumping 400,000 tokens of loosely relevant material produces worse answers than 20,000 tokens of the right material. Second, the window is a ceiling, not a target. Curating context is a skill, and it is the one that separates people who get remarkable results from people who get adequate ones.

What Claude costs: every plan compared

The API is separate and billed purely per token, with substantial discounts for batch processing and for prompt caching — the latter matters enormously if you send the same large context repeatedly, which most real applications do.

The decision most people get wrong is between Pro and Max. The honest test is simple: if you hit Pro's limits more than once a week and it interrupts real work, Max pays for itself immediately. If you hit them once a month, it does not.

The Claude interface, feature by feature

The chrome is deliberately quiet. What sits behind it:

• The composer accepts text, files, images, PDFs and pasted screenshots, and shows which model and effort level are active.
• Projects hold standing context: instructions, reference files, and every chat about one objective.
• Artifacts turn output into a persistent, editable side panel — documents, tables, diagrams, working web apps.
• Connectors and desktop extensions link Claude to your own systems, including Microsoft 365 on paid plans.
• Research runs a multi-step, multi-source investigation and returns a cited answer.
• Scheduled work runs a task on a recurring basis instead of waiting for you to ask.

Artifacts: things you can actually use

Most assistant output is text you copy somewhere else. Artifacts break that pattern: when Claude produces something substantial — a document, a table, a diagram, a small application — it appears in a panel beside the conversation, and you iterate on it in place.

This sounds like a UI nicety and behaves like a change of medium. Three consequences:

Iteration stops being destructive. "Make the second section shorter and add a risk column" edits the artifact rather than regenerating a wall of text you then have to re-read from the top.

Ideas become pokeable. Ask for an interactive calculator, a decision tree, a data explorer, and you get a working thing in seconds. For anyone who thinks by prototyping, this is the fastest path from vague idea to something you can react to.

Output becomes shareable. An artifact can be published as a link, which turns "here is what Claude told me" into "here is the tool."

The AI Hub covers the working habits in detail: build it in an artifact then iterate, iterate inside the artifact rather than re-pasting in chat, and turning an artifact into a link you can share.

Projects, memory and standing context

Anything you have pasted into Claude twice belongs somewhere permanent.

Projects are the main tool: a container with its own instructions, its own uploaded knowledge, and its own chats. A project for a client holds their brand guide and past work; a project for a codebase holds the architecture notes and conventions. The discipline that makes them work is one project per context, not one for everything — and pruning the knowledge as it goes stale, because outdated reference material is worse than none.

Memory carries facts across conversations. It is helpful and, like every such feature, it drifts. Check it periodically.

Custom instructions and style settings shape tone and format globally, which saves you restating "no preamble, British spelling, tables not paragraphs" forever.

Claude Code: the agent that lives in your terminal

Claude Code is the reason a large share of Claude's user base is developers. It is a command-line agent that reads your repository, plans a change, edits files, runs the tests, reads the failures and iterates — with you approving as it goes.

Four habits separate people who get real value from people who get a mess:

1. Let it explore before it edits. A first instruction of "read the relevant files and tell me how this works" costs one turn and prevents an edit built on a wrong mental model.
2. Approve the plan, not the diff. Reviewing intent is far cheaper than reviewing 400 lines of consequence.
3. Let it close the loop. Give it the test command. An agent that can run the tests and read the failure fixes its own mistakes; one that cannot hands you plausible code.
4. Keep a project instructions file that earns its place. Conventions, commands, gotchas. If a line in it never changes the output, delete it.

The Claude Code track in the AI Hub covers each of these as a standalone lesson.

MCP: connecting Claude to your own tools

The Model Context Protocol is an open standard, published by Anthropic and now widely adopted, for connecting an assistant to external tools and data. Instead of every app inventing its own integration, a tool exposes an MCP server and any MCP-aware assistant can use it.

In practice this means Claude can read from your issue tracker, query a database, search your documentation, or call an internal service — with the scope you grant and nothing more.

The failure mode is enthusiasm. Connecting eight servers on day one produces a model with a confusing surface and an access footprint nobody has reviewed. Start with one that earns its place and scope it to least privilege.

Constitutional AI: why Claude behaves the way it does

Every assistant is shaped after pretraining by an alignment stage. Most use reinforcement learning from human feedback, where raters pick the better of two responses and the model learns the preference. Constitutional AI adds a step: the model is given an explicit written set of principles and trained to critique and revise its own answers against them.

The observable results, in daily use:

• It disagrees more readily. Ask it to check your reasoning and it will more often tell you the reasoning is wrong, rather than finding a way to agree.
• It flags uncertainty rather than smoothing over it. "I am not confident about this specific figure" appears more often than in competitors — which is a feature, not a hedge.
• Its refusals have stated reasons. You can argue with a stated reason; you cannot argue with a vibe.
• It is steerable. Because the default behaviour comes from principles rather than accumulated preference, telling it what you want tends to move it rather than fight it.

This is not a claim that Claude is safe or correct. It hallucinates, it can be wrong with confidence, and it can be talked into things. It is a claim that its behaviour is more legible than most, which is a genuinely useful property in a tool you rely on.

Who uses Claude, and where

Anthropic publishes no country-level user breakdown, so the honest picture comes from web-analytics panels and from what the company says about its business mix. Both point the same way: Claude is far more concentrated than ChatGPT, both geographically and by use case.

That last point matters more than the map. A very large share of Claude's real-world use never touches claude.ai: it happens inside other companies' products, in terminals, and in enterprise systems through the API. Any comparison of "AI assistant market share" based on website visits systematically understates Claude for exactly that reason.

What Claude is genuinely good at

• Long-document work. Reading a contract, a codebase, a research corpus or a year of notes and reasoning across all of it at once. The million-token window is a real capability, not a spec.
• Writing that survives editing. Claude's default prose is less padded than most, and it responds unusually well to a voice guide. It is one of the few models where "match this style" produces something you would actually publish.
• Agentic coding. Multi-file changes, refactors, test-writing and debugging, especially through Claude Code. This is the use case Anthropic optimises hardest for and it shows.
• Analysis with its uncertainty attached. Ask for a recommendation and you tend to get the trade-offs, the assumptions, and an explicit note about what would change the answer.
• Being argued with. As a thinking partner it is materially better than the alternatives at taking the other side seriously instead of validating you.
• Structured output you can trust. For extraction and transformation into JSON or tables, it is reliable enough to build on.

Where Claude falls short

• It still hallucinates. Less floridly than some, but specifics — names, figures, citations, API signatures — need verifying. The careful tone is not evidence of accuracy.
• The knowledge cutoff is real. Each model's reliable knowledge stops at a stated date. Without search or supplied documents, recent events are invisible to it.
• Usage limits bite. On Pro, sustained heavy use — long documents, agentic coding — hits caps faster than people expect. This is the single most common complaint, and it is the reason the Max tiers exist.
• A smaller consumer ecosystem. Fewer consumer integrations, fewer third-party plugins, less presence in the tools your less technical colleagues already use.
• Over-thoughtful on simple tasks. Adaptive thinking sometimes spends real time on a request that wanted a one-line answer. Telling it to be brief works.
• Caution can read as friction. The same disposition that makes it flag uncertainty occasionally makes it hedge where you wanted a decision. Ask it to commit and it will.

The anatomy of a Claude prompt

Claude responds to the same fundamentals as every model, with two emphases that matter more here than elsewhere.

The second emphasis: say what to do, not only what to avoid. "Do not be verbose" gives the model a constraint but no target. "Answer in under 120 words, lead with the recommendation" gives it both.

The Claude prompt library

Twelve prompts built around what Claude is actually best at. Copy them, replace the bracketed parts, attach what they refer to.

Long documents and analysis

Writing

Code

The AI Hub keeps a larger, browsable prompt library organised by tool and task, with a lesson behind each move.

Seven advanced moves that change the output

1. Ask it to restate the task first. "Tell me what you think I am asking for and what a great answer looks like, before you start." One cheap turn that catches every misread.
2. Curate the context; do not dump it. With a million-token window the temptation is to paste everything. Relevance beats volume every time.
3. Give it a memory file for long projects. A living document of decisions, constraints and dead ends, kept in the project, so a fresh session starts where the last one ended.
4. Fan out, then merge. Independent sub-questions in parallel chats, then one chat to reconcile them, beats one thread trying to hold everything.
5. Add a check step between chain links. When one output feeds the next, insert an explicit verification turn. Errors compound otherwise.
6. Make it argue the side you did not pick. The single most useful thing Claude does that a search engine cannot.
7. Don't outsource the final judgement. Use it to widen the option space and pressure-test the reasoning. The decision stays yours, and so does the responsibility for it.

Claude vs ChatGPT vs Gemini

Benchmarks settle nothing that matters to you. Run your three most common real tasks through two of them for a week; the answer is usually obvious by Friday. The other half of this comparison is in our companion guide, what ChatGPT is and how to use it.

Privacy, data, and what Anthropic does with your work

• Commercial use is excluded from training by default. API, Team and Enterprise traffic is not used to train models unless you opt in. This is the clearest argument for putting a team on a paid business plan rather than letting people expense individual subscriptions.
• Consumer plans have a training setting. Free, Pro and Max users have a control over whether conversations may be used to improve models. Find it and set it deliberately.
• Enterprise controls are substantial. SCIM provisioning, audit logs, role-based access, spend controls, a HIPAA-ready configuration and a compliance API.
• Retention is finite but not instant. Deleted conversations are removed on a schedule, and safety systems may retain material longer in specific circumstances.

The organisational advice is the same as for any assistant: write one page saying what may never be pasted into an AI tool, provide a sanctioned option, and make it the easy path. People route around policies that make their work harder.

What Claude means for your visibility

When Claude answers a question about your industry, it does one of two things: it generates from what it learned in training, or it searches the web, reads a handful of pages, and synthesises an answer citing them. Increasingly it is the second.

That is the same shift happening across AI Overviews in Google, ChatGPT's search, and every answer engine that has followed. The consequence is blunt: if your page is not among the sources it reads, you are not in the answer at all. There is no page two to climb from.

What makes a page easy for a model to cite turns out to be what makes it good for a human: a direct answer near the top, self-contained factual statements that survive being lifted out of context, clean structure with descriptive headings and tables, and credibility corroborated elsewhere on the web. Structured data, semantic HTML and an llms.txt file all help a model parse and trust you.

The discipline is generative engine optimization. Compare it with classic search in GEO vs SEO, follow the playbook in how to get your brand cited by AI, or see how we run it as a capability.

Learn Claude properly: the free course

The Troiana AI Hub is a free, structured course in Claude — eight tracks of short, specific lessons, each built around one habit worth adopting. No sign-up, no email.

When you are done, take the Claude master quiz, or go sideways into the ChatGPT, Gemini and Cursor tracks.

A short Claude glossary

• Large language model (LLM) — the AI system underneath Claude, trained on vast text to predict and generate language.
• Haiku, Sonnet, Opus, Fable — Claude's model sizes, from fastest to most capable.
• Token — the chunk of text a model reads and writes, roughly three-quarters of a word. Context limits and API bills are counted in tokens.
• Context window — the maximum tokens a model can hold at once; one million on the current Claude line.
• Adaptive thinking — the model deciding how much internal deliberation a request deserves.
• Effort — the low / medium / high cap on how much of that deliberation it may spend.
• Artifact — a persistent, editable panel holding a document, diagram or app Claude produced.
• Project — a scoped workspace with its own instructions, knowledge files and chats.
• Claude Code — the terminal-native agent that reads a repository, edits files and runs tests.
• MCP (Model Context Protocol) — the open standard for connecting an assistant to external tools and data.
• Constitutional AI — Anthropic's alignment method, training a model to critique its own output against written principles.
• Responsible Scaling Policy — Anthropic's published framework of capability thresholds and required safeguards.
• Hallucination — a confident but false statement produced by a model.
• Retrieval-augmented generation (RAG) — grounding a model in retrieved documents so it can cite real sources.
• Generative engine optimization (GEO) — structuring content so AI answer engines find, trust and cite it.

Common questions