Anthropic: the company behind Claude.

What does Anthropic do?

Anthropic develops generative AI: systems that can produce text, analysis or code from input. Its work includes model training, research, evaluation and product development. Current Claude models can also process visual input. What a user can actually do additionally depends on the software and tools surrounding the model.

A business can work with this technology in several ways. An employee discusses a quotation in the Claude app. A developer uses the API to classify customer questions. A programmer uses Claude Code to investigate a bug in an existing application. The Anthropic name identifies the technology provider; it does not automatically explain which data a particular application may read or which actions it may perform.

For marketing teams, the combination of language, documents and software is particularly relevant. A model can turn product information into a first draft, identify differences between versions or explain a table. The organisation remains responsible for the accuracy of its offer and the final publication.

Research into capabilities and behaviour

Anthropic studies both what models can do and how their behaviour can become more understandable and controllable. This includes interpretability research, evaluations and safeguards. Interpretability seeks to understand what happens inside a model; a convincing explanation generated by the model itself is not automatically evidence of that process.

Public model and system cards describe tests, limitations and measures. Read them alongside current product documentation. A laboratory test with a particular setup does not directly predict how a model will handle your untidy documents, exceptions or software environment.

A practical evaluation can start small. Give two models the same anonymised customer questions and assess correct classification, missing information, unwanted assumptions and necessary corrections. Record the model version and settings. This produces more useful evidence for your organisation than simply adopting a general leaderboard.

What is Constitutional AI?

Constitutional AI is an approach described by Anthropic for guiding model behaviour through explicit principles. The research published on 15 December 2022 describes self-critique and revision of answers, followed by learning with AI feedback. People choose the principles and setup; the system does not independently invent a socially binding constitution.

On 22 January 2026, Anthropic published a new, more detailed constitution for Claude. It describes intended behaviour and the judgements the company wants the model to make. A normative document explains the developer’s aims. It does not establish that every individual response meets them.

Distinguish behavioural policy, technical safeguards and factual reliability. A model can politely refuse certain requests and still make a mistake in an ordinary summary. Businesses still need source checks, clear task boundaries and appropriate access to information.

Claude, the API and Claude Code

Claude is the assistant people interact with in natural language. The same name also covers models such as Sonnet, Opus, Fable, Mythos and Haiku. A new model version and a new app feature are different changes. Features also do not necessarily appear in every plan or channel simultaneously.

The API supports software integration. Your application sends instructions and data to a chosen model and processes the response. An online shop might generate proposed product categories, then have rules or an employee check them. API usage has its own billing, limits and model access; an assistant subscription does not automatically provide an unlimited API budget.

Claude Code is a system for programming work. Current documentation describes use in environments including the terminal, IDE, desktop and browser. It can explore a repository, propose or make changes, and use commands and checks within its permissions. Claude Code is the product; the Claude model inside it provides part of the reasoning and generation.

Tools and agents

An AI agent combines a model with a goal, context and tools. The model can propose a next step, but the application determines whether a file, search function, terminal or external service is actually available. Anthropic provides models and developer support for this. A model alone does not become an independently operating employee.

With computer use, a system can analyse screenshots and interact with an interface through tools. The environment executes the actions. Connecting a system does not grant general permission for every possible change. An agent preparing a campaign might gather information and complete a draft while publication remains a separate step.

In business automation, make the sources used and changes made visible. A useful result includes verifiable changes and a clear completion point, as well as an answer. This matters particularly when one task involves several systems.

Business products and purchasing decisions

Anthropic also offers business Claude products, including Team and Enterprise. Alongside model capabilities, organisations should examine administration, access controls, data arrangements and available integrations. The current pricing page is the source for the offer; amounts and usage limits can vary by product and contract.

Compare the cost of a completed task. A fast model that performs straightforward classification correctly may suit a high volume of repetitions. A more capable model may save correction time on a difficult analysis. Extra capacity adds little value when the required source is missing or nobody checks the output.

For example, an agency wants to prepare monthly reports. Start with one report, record where each figure comes from and review the explanation. This reveals the integration, model and permissions the workflow actually needs. A trial should test your working process, not merely produce an appealing demonstration.

A short timeline

How should you read current model information?

All five families remain relevant on the review date. Their latest versions differ: Sonnet 5.5, Opus 5.5, Fable 5.1, Mythos 5.1 and Haiku 5.5. Mythos requires verification. One family’s number is therefore not a reliable indication of another family’s current version or availability.

Before implementation, check three sources: the current model overview for specifications, lifecycle documentation for retirement, and the relevant product documentation for access. A historical introduction remains useful for understanding development, but does not replace these sources. Read the individual entries for practical differences and examples.