Sonith ManakkatIf you have heard people say "ask Claude" and wondered what they mean, this guide is for you. No technical background is needed. By the end you will know what Claude is, who makes it, what it is good at, where it goes wrong, and how to try it safely. Each section gives you the plain-language picture first, then an optional Expert view for readers who want the precise details.
The short answer
Claude is an AI assistant. You type (or paste, or upload) something, and it writes back. You can ask it to explain an idea, draft an email, summarise a long document, look at a picture, or help write computer code. It is made by a company called Anthropic.
Expert viewAnthropic and its stated mission
Claude is made by Anthropic, which describes itself as a public benefit corporation. Its company page states its mission as "AI research and products that put safety at the frontier". Its governance section mentions a Long-Term Benefit Trust and a board elected by stockholders. Those are the company's own statements about itself, so read them as intent, not as a guarantee of any particular outcome.
This guide is an independent article. I have no connection with Anthropic, and every fact below is linked to an official page at the end.
What is a language model?
Claude belongs to a kind of software called a large language model. The idea is easier than the name. Imagine your phone suggesting the next word as you type. Now imagine that suggestion engine trained on a huge library of books, articles, websites and computer code, and made far more capable. Given some text, it works out what text should come next, then the next piece, and so on. A reply is built one small piece at a time.
- 1Your text goes inYour words, plus any file or picture, become the prompt.
- 2It is split into tokensTokens are small pieces of text, such as a short word or part of a longer one.
- 3The model scores what comes nextIt works out a probability for each possible next token.
- 4One token is pickedPicked from those probabilities, which is why wording can differ between runs.
- 5The token is added and the loop repeatsThe longer text goes back in and the next token is predicted, thousands of times.
- 6The reply is completeIt was built one small piece at a time, with no step that checks facts.
Takeaway. A reply is a chain of likely next pieces, so fluent does not mean verified.
That simple loop turns out to be powerful. To continue text well, a model has to pick up grammar, facts, styles and patterns of reasoning. That is why one tool can help with a poem, a spreadsheet formula and a polite complaint letter.
Claude looks things up like a search engine
By default it writes from patterns it learned in training. Some products add a web search tool, but that is an add-on.
Claude knows me and remembers our past chats
It mostly knows what is on the page in front of it. Some products add a memory feature, so check what yours does.
Expert viewHow the "next piece" works
Models do not read letters or whole words. They read tokens, small pieces of text, and predict a probability for each possible next token, then pick one. Repeating that thousands of times produces a reply. Because the output is built from probabilities, the same question can produce differently worded answers, and a fluent answer is not the same thing as a verified one.
Modern assistants are trained further after the first stage so that they follow instructions and behave helpfully and safely. Anthropic describes its alignment, interpretability and Frontier Red Team work on its research pages. For you as a user, the practical lesson is the same: treat the output as a strong draft, not a final authority.
The Claude family, as of 21 October 2026
Anthropic does not make one Claude. It makes a family of models of different sizes, a bit like a range of cars. Its documentation lists four current models, and says all of them accept text and images and can use tools.
- Claude Haiku 5.5The fastest. For high-volume jobs such as sorting and extracting information.
- Claude Sonnet 5.5The best combination of speed and intelligence, according to Anthropic.
- Claude Opus 5.5For long-running coding and knowledge work. Anthropic suggests starting here for most workloads.
- Claude Fable 5.1For demanding reasoning and long, multi-step agentic work. The slowest of the four.
According to Anthropic's news page, Fable 5.1 was announced on 15 September 2026, Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. Anthropic also mentions Claude Mythos models with limited availability; I have not covered those. Lineups change often, so treat this as a snapshot and check the official page.
- 4current models
- 1Mtokens of context window
- Jun 2026reliable knowledge cutoff
Expert viewReading the model table
Anthropic's overview lists a 1M-token context window for all four current models, a maximum output of 128K tokens per request, adaptive thinking, and a "reliable knowledge cutoff" of June 2026. The docs say 1M tokens is roughly 555,000 words on the current tokenizer. A knowledge cutoff is the date after which the model has no built-in knowledge, so anything newer must be supplied or searched.
Model IDs are pinned snapshots. The IDs follow the pattern claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5 and claude-fable-5-1. Anthropic states retirement commitments (not sooner than September to October 2027 for these four on platforms it operates), which is useful when you build something that depends on a model. Prices are covered in part 4 of this series.
What Claude can do
- Explain and teachAsk for any idea at the level you choose: "explain this to a ten-year-old".
- Write and editEmails, reports, summaries, rewrites in a different tone.
- Read long materialAnthropic says the window holds roughly 555,000 words.
- Look at imagesPhotos of documents, charts or screenshots can be described and discussed.
- Help with codeWriting, explaining and fixing software is a well-known use.
- Work with toolsConnected to other software, Claude can search, calculate or act on files.
- You to Claude You ask for something that needs more than text, such as a calculation or a search.
- Claude to A tool Claude requests the tool, for example web search or running code.
- A tool to Claude (reply) The tool's result is returned into the conversation.
- Claude to You (reply) Claude carries on from the result and writes the reply.
Takeaway. Tools add abilities to text-only Claude, and their results become part of what it reads.
What Claude cannot do (or does badly)
Claude is good at
- First drafts, summaries, explanations, reformatting
- Spotting patterns and gaps in text you give it
- Working fast on routine, well-defined tasks
You must still do
- Checking facts, numbers and quotes that matter
- Judgement, ethics and accountability for decisions
- Professional advice on health, law or money
- It can be wrong. Models sometimes state invented facts, names or numbers with complete confidence. People call this "hallucination".
- Its memory is limited. What it knows about your conversation is mostly what is on the page. Some products offer a memory feature.
- It cannot see the future. Its reliable knowledge stops in June 2026 unless it is given newer information or can search.
- 1Would money, health, law or safety depend on it?Yes Check a trusted source or a professional before acting.No keep going
- 2Does it contain a fact, number, quote or name you will repeat?Yes Check it against a source. Ask Claude for the quote it relied on.No keep going
- 3Is it a summary or rewrite of text you supplied?Yes Skim it against your original.No keep going
- 4OtherwiseA draft or idea: read it, then edit it into your own words.
Takeaway. The higher the stakes, the more you check.
Expert viewWhy errors are structural, not rare accidents
A model optimises for plausible continuations. When the training data is thin, or the question has no clear answer, a plausible-sounding but wrong continuation is still a high-probability one. That is why the failure is "confident": nothing in the output marks which sentences are shaky.
Anthropic says its latest models are less prone to hallucination than earlier ones and publishes prompting guidance such as asking the model to read the relevant files before answering about them. It also says, in the Opus 5.5 announcement, that the model often suspects it is being evaluated, which limits how well testing predicts real-world behaviour. Good practice follows from this: ask for quotes and sources, ask the model to say when it is unsure, and verify anything you will act on.
Claude Code and agents
Beyond chat, there is Claude Code, which Anthropic describes as an agentic coding tool that reads your codebase, edits files, runs commands and works with development tools. It runs in a terminal, in code editors, as a desktop app and in a browser.
An agent takes steps toward a goal instead of answering once: it looks, plans, acts and checks.
Anthropic's documentation says Claude Code can run several agents in parallel, with a lead agent coordinating. Later parts of this series cover context, CLAUDE.md, skills, sharing with a team, and agents in depth.
How Anthropic describes its approach to safety
Anthropic's research pages name teams working on alignment (helping models be helpful, honest and harmless), interpretability (studying how models work inside) and a Frontier Red Team that stress-tests systems. For Opus 5.5 the company says external evaluators tested the model before release, that it is more resistant to prompt injection than its predecessor, and that it is less likely to take hard-to-reverse actions.
- Keep secrets outNo passwords, bank details or other people's private data in prompts.
- Check the settingsPrivacy and data options differ between personal and business plans and can change.
- Verify what mattersCompare important answers with a trusted source.
How to start
- Open claude.ai and try a small, low-stakes task, such as summarising an article you already know.
- Be specific about what you want, and say who the answer is for.
- Check the result against what you know. Ask follow-up questions when something looks off.
- Move on to part 2 of this series for the basics that make answers better.
Frequently asked questions
Is Claude the same as ChatGPT?
No. They are different products from different companies, though both are language models and feel similar to use.
Is it free?
Anthropic offers a free plan and paid plans. Part 4 explains plans and pricing, with the official page linked.
Does Claude think?
It produces useful reasoning-like text, but whether that is thinking is a debate. In practice, treat it as a capable tool, not a person.
Can I trust it?
Trust it as you would a quick, clever colleague: useful, but verify what matters.