Claude: what this AI is and how to use it

Claude is Anthropic's family of AI models and the chat interface built around them: you describe a task in plain words, and the model reads documents, writes text and code, and explains someone else's spreadsheets and reports. There isn't one model inside but several — from fast and cheap to the most capable — and you choose between them based on the task.

Below: what these models are and how they differ, what they cost, how to get started, how access works from Russia, and the point that matters most for a company — what happens to the data an employee sends into the chat. Russian-language reviews usually skip that last question, and it is the one that decides whether the tool can be let anywhere near your working environment.

What is Claude, in plain terms?

In short: a chat with an AI model that can work with your own files and hold the context of a large project.

The official definition from the documentation:

"Claude is a family of state-of-the-art large language models developed by Anthropic" — the Claude platform documentation.

The practical difference from ordinary search is that Claude works with the material you give it. Upload a forty-page contract and it will find the termination clauses. Give it an export from your accounting system and it will build a summary. Paste in a thread with a contractor and it will pull out the commitments and deadlines.

You can use it through the claude.ai website, desktop apps for macOS and Windows, mobile apps, and — for developers — through the API and the Amazon, Google and Microsoft clouds.

If this is your first encounter with the idea of assigning work tasks to an AI in plain language, start with the companion piece — what vibe coding is and why a business should care. That one is about the approach; this one is about the specific tool.

Which models are in the family, and how do they differ?

In short: four models for different tasks and budgets; the more capable, the more expensive and the slower.

The line-up as of August 19, 2026, according to Anthropic's documentation:

Here is what matters if you don't write code.

A token is a chunk of text roughly three or four characters long. A million tokens of context, as the documentation explains, corresponds to about 555 thousand words. For comparison: War and Peace runs to around 500 thousand. In other words, the model can hold a small library of contracts in view at once.

The prices in the table apply to API access, where you pay for actual usage. On a regular subscription you pay a fixed monthly amount and switch between models with a selector.

The documentation itself recommends starting with Opus 5 for complex work tasks, and reaching for Fable 5 where you need maximum capability. Haiku makes sense when speed and volume matter more: bulk processing of similar requests.

There is a fifth model as well — Mythos 5 — but it is handed out by invitation only, for defensive cybersecurity work; you cannot sign up for it yourself.

What can Claude actually do?

In short: read files, write and rewrite text, work with code, assemble interactive documents, and connect to your working services.

The product's stated capabilities, ordered by practical value to a company:

• Working with documents. PDF, Word, Excel, images. Reading what you upload: find the relevant part, compare versions, produce a digest.
• Projects. A workspace where context carries over between sessions. Upload your policies and templates once and the model remembers them across every conversation inside that project.
• Artifacts. A result you can use straight away: a table, a document, a small calculator, a diagram.
• Code. Debugging, review with explanations, optimization. For serious work there is a separate product, covered below.
• Integrations. Google Workspace (mail, calendar, documents), JIRA, Zapier, plus your own tools connected through the open MCP protocol.
• Voice mode and a browser extension — for when speaking is faster than typing.

Worth naming separately is Cowork — a mode where tasks are delegated on a schedule instead of being handed over one at a time by hand.

The volume of material the model can hold at once deserves its own paragraph: it is what separates a working tool from a toy. A million tokens of context on the larger models is a year of project correspondence, a company's full set of policies, or several dozen contracts loaded at the same time. The model sees all of it and can answer questions that require comparing different documents against each other: where the terms diverge, what changed between versions, which obligations overlap.

How do you start using Claude?

In short: sign up at claude.ai, pick a model, upload your files and describe the task in ordinary words.

1. Open claude.ai and create an account. During sign-up you will be asked about using your data for training — that choice matters, and we come back to it in its own section.
2. Start on the free plan: it tells you whether the tool suits you before you pay anything.
3. Describe the task the way you would explain it to a new hire: what the material is, what you need out of it, in what form. Vague briefs produce vague answers.
4. Upload files straight into the chat. Working on your actual document is almost always more useful than a conversation in the abstract.
5. Set up a Project for a recurring task and put the permanent material in it — policies, templates, examples of answers that worked.
6. Check the result on the substance. The model states wrong claims just as confidently as right ones.

What does Claude cost, and is there a free plan?

In short: there is a free plan, an individual subscription runs 17–20 dollars a month, and a team seat starts at 20 dollars.

Plans according to the official page as of August 19, 2026:

A practical guide to choosing. The free plan is enough to try it out and decide whether this is your kind of tool. Pro covers one person's daily work. Team plans are worth it less for the limits than for control: shared billing instead of ten personal cards, single sign-on, and an administrator who can set data-handling rules for everyone at once.

What about access to Claude from Russia?

In short: officially the service is unavailable in Russia — and that is a vendor restriction.

Russia is not on Anthropic's list of supported countries, neither for Claude.ai nor for commercial API access. The list contains 195 countries and territories. Belarus is not on it either, and Ukraine carries a carve-out for Crimea and four regions.

The company's own wording is that Anthropic reserves the right not to provide products and services to organizations whose direct or indirect ownership traces to countries outside that list.

Two practical conclusions follow, and both matter more than the question of "how do I get in."

First: a company building a workflow on this tool has no contractual relationship with the supplier. Which means no predictability either — access can disappear at any moment, and nobody will have breached anything.

Second: even once access is solved technically, there remains a question no setting closes — the data goes to servers outside Russia. A draft email doesn't care. A customer register or a financial model, though, runs straight into the question of cross-border transfers and Federal Law 152-FZ, Russia's personal data protection law.

What happens to the data you send?

In short: on commercial plans they do not train on your data by default; on personal plans it depends on what you chose at sign-up, and the retention period differs by a factor of sixty.

This is the most important section if you are thinking about a company rather than personal experiments. Let's go through Anthropic's official policy.

Commercial products — API access and Claude for Work:

"By default, we will not use your inputs or outputs from our commercial products to train our models."

There is one exception: if a user clicks a feedback button to rate an answer, that conversation may go into training — it is kept for up to five years and detached from the user identifier. An organization's administrator can turn the feedback buttons off in the privacy settings.

Consumer plans — Free, Pro and Max: the user makes the choice at sign-up. The consequences differ substantially:

The setting can be changed at any time in the privacy preferences.

What this means for a manager. If five employees have taken out personal subscriptions and are uploading work documents into them, the way those documents are handled is governed by a checkbox each of them ticked when signing up a year ago and almost certainly does not remember. A corporate plan solves exactly that problem: the rules are set centrally and identically for everyone.

There is a separate layer that no plan resolves. The data physically leaves the company's perimeter and the country. Drawing the line between what may go outside and what must stay in is architectural work, and we do it in custom projects from the outset, not after an incident.

Claude and code: how is Claude Code different from Cursor?

In short: the chat helps with code in words, Claude Code carries out tasks in your project itself, and Cursor is an editor where the same model works inside a familiar interface.

Three different tools that are easy to confuse:

• The regular Claude chat. You paste in a fragment of code, get an explanation or a fix, and paste it back by hand. Good for making sense of someone else's script or repairing a formula.
• Claude Code. A separate product: the agent opens the whole project, edits files, runs commands and reads the output. We have a detailed piece on how it works and which working habits pay off — Claude Code: what it is, how it works and how to use it.
• Cursor. A code editor where the agent lives inside the interface developers already know, and you can pick the model, Claude included. A breakdown with pricing and limitations — Cursor AI: what it is, how to use it and what it costs.

For a non-technical manager the difference comes down to one thing: the chat saves an employee time, while the other two change how software comes into existence at your company. The second calls for a different level of oversight.

Where Claude gets it wrong, and what not to expect

In short: the model phrases things confidently, including when it is wrong, and it does not know what isn't in its data.

Three limits worth keeping in mind from day one.

Knowledge stops at the training date. Each model has a point up to which its knowledge is reliable: for Opus 5 that is May 2026, for Sonnet 5 and Fable 5 January 2026. Anything later the model either does not know or reconstructs by analogy. Current prices, laws and exchange rates you have to supply yourself.

Confidence is not correctness. A wrong answer looks exactly as convincing as a right one: the same smooth phrasing, the same readiness to go into detail. You have to check on substance, not on tone.

Responsibility cannot be delegated. The model is not accountable for the consequences of its answers — neither legally nor practically. If a delivery decision or a settlement with a counterparty rests on its output, the person who made that decision is the one who answers for it.

Hence a simple working rule: the more expensive the mistake, the tighter the check. On a draft email you can trust it almost blindly; on a contract, reread the whole thing.

How companies use Claude at work

In short: wherever there is a lot of text and routine reading, and where a mistake is cheap.

The scenarios that pay back fastest:

• Processing incoming documents. Contracts, invoices, proposals: pull out the terms, compare against the previous version, assemble a summary for a decision.
• Drafts and correspondence. Replies to clients, internal policies, job descriptions, instructions — a first version instead of a blank page.
• Working with data. Explain an export, find anomalies in a table, formulate a query against the accounting system.
• Internal knowledge. A project loaded with company policies turns into a reference desk that answers new hires instead of their colleagues.
• Analysis and meeting prep. Condense reports into a short memo, draw up questions for a contractor.

From our own custom development practice a pattern stands out: the rollouts that win are the ones where AI was placed on a specific step with a clear input and output — say, reading incoming source documents before they are entered into the accounting system. The ones that fail are attempts to "adopt AI" in general, with no answer to the question of whose work it removes and at which step.

What the successful scenarios have in common: the result is checked by a person who knows the subject, and the cost of a mistake is measured in time spent. The moment the result travels further unchecked and touches money, you need a different level of control — with logic, access rights and reconciliation, which is to say a system, not a chat. There is a separate piece on that class of software — AI agents for business: what they are and which kinds exist.

Who is Claude enough for, and who needs their own system?

In short: as an employee's working tool — almost everyone; as the foundation of a process that money depends on — no one.

Claude covers three jobs well. Taking the routine of reading text and documents off people. Speeding up the preparation of drafts. Giving quick access to knowledge sitting in files and in people's heads. In all three the decision is made by a person, and the model saves their time.

The job it does not cover, let's name plainly: becoming the place where a company's accounting, obligations and money live. The reason is a set of process requirements a chat simply doesn't have: access rights, change history, data reconciliation, predictable behavior when an external service goes down, and a contract under which someone is answerable for the outcome. We covered the fork between a quick fix and a system built around your own process separately — your own system or an off-the-shelf box.

At INCUBE AI we use these models daily and see the boundary without illusions: where AI saves weeks, we put it in; where a system is needed, we build a system. If you are sizing the tool up against your own tasks — come in for a consultation: we will look at the process and tell you straight which parts a 20-dollar subscription covers and which need to be built as an engineering system — under contract, with the data held in Russia and with support after handover.

Sources

• Claude product overview — official site
• Model family, pricing and context — platform documentation
• Claude pricing plans
• Anthropic's list of supported countries
• Is my data used for model training — Anthropic Privacy Center