How Claude actually works: a practical guide to Anthropic's AI assistant

01What Claude is and who builds it

Claude is a family of large language models developed by Anthropic, an American AI company headquartered in San Francisco. According to Wikipedia's summary, Anthropic was founded in 2021 by former members of OpenAI, including Dario and Daniela Amodei, now CEO and president, with the explicit goal of advancing AI safety, and Claude, its flagship product, first shipped as a chatbot in March 2023. The company is also closely tied to Amazon, whose massive Trainium-powered data centers provide much of its training compute, a story we covered in detail in our previous dispatch on Amazon's AI infrastructure.

In practice, "Claude" is three things at once. It is the underlying model family that other companies license and embed in their own products. It is the consumer chat interface at claude.ai, with free and paid tiers. And it is a set of developer tools, most notably the Claude Code agent, built on the same models through Anthropic's API. The same underlying model can behave quite differently across these surfaces, because the surrounding system, instructions, tools, and available data, differs.

The name shows up in one more place that confuses people: AWS's Bedrock and other cloud platforms resell Claude alongside competing models. Anthropic the company builds and controls the models; the clouds distribute them. That distinction matters when you are deciding where to run Claude and under whose terms.

02The LLM foundations: tokens, context windows, training

Underneath the branding, Claude is a large language model. Wikipedia's summary describes LLMs as AI models trained on vast amounts of text for natural language processing, especially language generation, and notes they are the basis of most modern chatbots, including ChatGPT, Claude, and Gemini. Everything the model does is, mechanically, predicting what text comes next. Nothing more exotic than that, done extraordinarily well, at enormous scale.

The unit of prediction is the token, a chunk of text that is usually a word, part of a word, or a piece of punctuation. You pay per token, you are limited by tokens, and "thinking" consumes tokens. English text averages roughly four characters per token, so a 1,500-word memo is on the order of 2,000 tokens. Once you internalize tokens, pricing tables and rate limits stop being mysterious.

The context window is the model's working memory: the total number of tokens, your conversation, pasted documents, instructions, and the model's own output combined, that it can consider at once. When the window fills, older content falls out of view. This is why a chatbot "forgets" the beginning of a very long conversation, not because of a bug, but because the math literally has nowhere to put it. Claude's context windows have grown enormously across generations, from roughly 9,000 tokens at launch to a million tokens in recent versions, a trajectory the chart below lays out.

Training happens in phases. Pretraining on a huge corpus of text and code teaches the model language, facts, and reasoning patterns, at a cost of months of time on tens of thousands of chips. Fine-tuning then shapes behavior, and a final alignment stage, which Anthropic has pioneered under the name constitutional AI, uses written principles and model feedback rather than only human labels to make the assistant helpful and less prone to producing harmful output.

03What makes Claude different from ChatGPT and Gemini

All three leading assistants are large language models with chat interfaces, coding abilities, and multimodal input. Wikipedia's LLM summary lists them side by side as the basis of modern chatbots, and at the level of "ask it a question, get an answer," they are interchangeable. The differences that matter in daily use are in emphasis, and they shift with every model release, so treat any comparison as a snapshot.

Claude's reputation rests on three pillars. First, long-form quality: writers and analysts consistently pick it for drafting and editing substantial documents, citing a more controlled tone and fewer cliches. Second, code: Claude has become the default model inside many coding tools and its benchmark presence in real-world software engineering tasks is strong. Third, long context: with windows up to a million tokens, Claude can genuinely hold a large codebase or a book-length document set in view, where shorter-window competitors must summarize or retrieve piecemeal.

OpenAI's ChatGPT retains the broadest ecosystem: plugins, enterprise adoption, image generation, and the largest user base, which makes it the safe default for general-purpose use. Google's Gemini is woven into Search, Docs, and Android, and its integration with Google's own ecosystem is unmatched for anyone living there. The honest summary is that professionals increasingly use all three, picking per task, and the chart below compares their publicly stated context windows as a concrete, if narrow, differentiator.

04Projects, artifacts, and the agentic workflow

Most people use Claude as a question-answer machine and stop there. The features that change productivity are the ones that give it persistent memory and structure. Projects, on paid plans, let you pin a set of documents and standing instructions so that every conversation in the project starts with that context already loaded, instead of you re-pasting your style guide or product spec each time. It is the difference between an assistant who has read your files and one who has not.

Artifacts are the other half. When Claude produces something structured, a document, a web page, a diagram, a small application, it can render it as a live, editable artifact beside the conversation rather than dumping it as text. You can iterate on it in place, and for code especially, the write-run-fix loop is dramatically faster than copy-pasting into an editor. A single well-scoped artifact request can replace an afternoon of tooling setup.

"Agentic" is the current buzzword tying it together. An agentic workflow means you give Claude a goal rather than a question, and it plans, uses tools, reads results, and corrects course over many steps. Telling it to research a topic and draft a brief, with web search and file tools enabled, produces a materially different class of output than asking for an essay in one shot, because the model can ground its claims in what it actually found. The skill on your side is task decomposition: stating goals, constraints, and what "done" looks like.

05Claude Code and developer workflows

Claude Code is Anthropic's agentic coding tool: instead of you asking for a snippet and pasting it into your editor, you point Claude at a repository, describe the change in plain language, and let it read the relevant files, edit them, run the tests, and iterate until they pass. It is the fullest expression of the model's strengths, long context over a real codebase, strong code generation, and tool use in a loop.

Used well, it shifts the developer's job from typing code to reviewing changes. That is faster for mechanical work: adding an error-handling path across a dozen call sites, writing the tedious test suite for a module, migrating a deprecated API, spelunking an unfamiliar corner of a large repo. Used badly, it produces plausible code that is subtly wrong, which is why the recommended pattern is small, verifiable steps with tests as the gate, rather than "build me the whole feature."

The ecosystem effect is worth noting. Claude models also power a large share of third-party coding assistants and IDE integrations, and Wikipedia's summary of Claude notes its use in AI-assisted software development alongside the chatbot product. For Anthropic, developers are the most commercially important user base, and the API plus Claude Code plus IDE integrations form a coherent stack aimed squarely at them.

06Strengths and known weaknesses

Where Claude consistently earns its keep: long documents and long context, where competitors with smaller windows must summarize or chunk; drafting and editing prose that sounds like a person wrote it; multi-step tasks with clear goals; and code, both one-off scripts and repository-scale changes. The constitutional AI alignment approach also shows up in daily use as a tendency to state uncertainty and refuse overconfident fabrication, which is a genuine quality, not a marketing one.

The weaknesses are the honest flip side. Hallucination is not solved anywhere in the field: ask any of these models for a fact it half-remembers, and it can produce a confident, specific, wrong answer, complete with a plausible citation. Long context does not mean perfect recall; models can miss details buried in the middle of a very large context. And each release is a fresh roll of the dice on tone and behavior, which is why teams that build workflows on a specific model version often pin it.

There are also structural limitations no model can escape: training data has a cutoff, so anything recent requires live tools like web search; the model has no memory of you between sessions except what you deliberately provide via Projects or pasted context; and it is genuinely bad at the things it was never going to be good at, like exact arithmetic on long numbers or counting, which it does approximately in tokens rather than exactly. Knowing the failure modes is half of using the tool well.

07Pricing tiers and what $20 actually gets you

The free tier at claude.ai gives access to a current model with daily message limits, enough to evaluate whether the assistant fits your work. The Pro tier at $20 per month is the mainstream option: substantially higher usage limits, access to the more capable models in the family, Projects, file uploads, and early access to new features as they ship. For most individuals doing serious weekly work, this is the tier that changes productivity.

The Max tier, at roughly $100 and $200 per month depending on level, raises usage limits much further and is aimed at people running Claude Code or long agentic sessions all day, where Pro limits genuinely bind. Team and Enterprise plans add centralized billing, administration, and larger shared context, and API pricing is usage-based per million tokens, with input tokens cheaper than output tokens and prompt caching reducing the cost of repeated long context, which is how heavy users keep a large document loaded affordably.

A practical note on the economics: heavy chat users on Pro will hit limits during long agentic sessions, because a single tool-using task can consume a large context many times over. If you find yourself rationing messages, the answer is either Max or the API, and the crossover point is easy to estimate once you know your token consumption. Otherwise, $20 buys more capability than almost any software subscription in history, and the limit you will actually feel is usage volume, not intelligence.

08How to get the most out of it

First, set up Projects for anything recurring. If you write in a particular voice, keep documents to a template, or repeatedly answer questions about the same body of material, a Project with the reference files and instructions pinned turns every future conversation into one that starts informed. Five minutes of setup, repaid the first time you skip re-pasting a spec.

Second, feed it your real material. Claude's long context is its most distinctive capability, and the most common mistake is under-using it: describing a document instead of attaching it. Paste the contract, the code file, the transcript. Ask for a specific output with a stated audience and length. The difference in output quality between "summarize this" on a described document and on the actual text is not subtle.

Third, iterate in the conversation instead of restarting. Telling Claude what to change about its own draft, shorter, keep the structure, address this objection, is faster and better than re-prompting from scratch, because the context already contains its reasoning. Fourth, verify anything factual that matters. Treat confident specifics, numbers, quotes, citations, as a first draft to check, not a lookup result. Used this way, as a drafting and synthesis engine with a human in the verification loop, Claude is one of the highest-leverage tools a knowledge worker can currently have. The people getting the most out of it are not the ones with the cleverest prompts; they are the ones who bring it their real work.