Claude in one minute
Use Claude to understand a difficult letter, compare documents or explain code. Current information requires suitable sources or search tools. Taking actions requires connected tools and permissions. Model choice affects capabilities, speed and cost, but does not replace a good brief and checking the result.
Claude, Anthropic and an LLM
Anthropic develops the technology. Claude is the brand used for the assistant and models. An LLM is the broader technical category: a large language model that uses learned patterns to process input and generate output. Not every LLM is Claude, and an application using Claude need not resemble the familiar chat interface.
The name shown in a model picker identifies an option available within that product. The API has its own model IDs and conditions. A statement such as “Claude has one million tokens of context” is therefore too broad. The selected model, application and setup determine what a particular task can use.
Anthropic introduced Claude on 14 March 2023. The Claude 3 family made Opus, Sonnet and Haiku recognisable choices in 2024. Fable and Mythos joined the offering in 2026. This history describes how the range developed; the current catalogue needs a separate check.
Writing, explaining and comparing documents
Claude can draft, restructure, translate and explain complicated material in simpler language. A clear instruction includes the goal, audience, available facts and desired format. For a product description, ask it to use only features from the supplied catalogue and flag missing information.
Document work extends beyond a short summary. You can compare two versions, collect obligations or organise customer interviews by theme. Request references to the original passages. This allows you to check whether the model actually bases a conclusion on the material.
Images can also serve as input, such as a screenshot or chart. The current models in this comparison support text and image input and produce text. This does not mean that every Claude model generates images or audio itself. File processing and additional media features may be provided by the application surrounding the model.
An everyday example: upload two versions of a quotation and ask which amounts, delivery dates and conditions changed. Treat the comparison as a draft check and reread the relevant passages yourself. A well-written answer can still miss a line.
Research and current sources
A model response is not automatically a live search. Where the Claude environment provides web search, research or connected sources, these can supply current information. The features visible to you depend on the product, account and permissions. For time-sensitive questions, check that the sources were actually consulted.
Break a research question into smaller questions: what is known, what is missing and which source could answer it? Ask for publication dates and a distinction between source information and interpretation. A reference must support the associated claim; the existence of a URL alone does not prove that it does.
A marketing agency might use Claude to compare three supplied research reports. A useful result identifies differences in audience, measurement period and method. A single average without that context could mislead. The aim is a verifiable analysis, not the longest possible summary.
Long context: extensive material, not perfect memory
A context window is the amount of information a model can consider during processing. It may include instructions, conversation history, documents and tool results. Tokens are processing units for text and other input; they do not equal a fixed number of words or pages.
On 7 October 2026, API documentation lists a 1 million token context window for Sonnet 5.5, Opus 5.5, Fable 5.1 and Mythos 5.1. Haiku 5.5, introduced on 7 October 2026, also has 1 million tokens; the older, still active Haiku 4.5 has 200,000 tokens. These are API model specifications, not a promise that a Claude subscription accepts the same volume of files or conversation text in every situation.
Having more space does not guarantee reliable retrieval of every detail. Organise sources, identify document versions and ask focused questions. Keep important decisions outside the conversation too. Context also differs from persistent memory: information from a previous task is not automatically available in the next one.
For a large knowledge base, RAG, which retrieves relevant passages selectively, may be more useful than resending every document. Choose an approach based on source discoverability, freshness and verifiability.
Programming and Claude Code
Claude can explain code, help locate a bug and propose changes. Actual project work depends on the repository, execution environment and tools. Without execution, the model cannot demonstrate that a proposed change passes the tests.
Claude Code is Anthropic’s product for this type of programming work. It can inspect and edit files, use commands and run checks within its permissions. The product organises the workflow; a selected Claude model supplies the underlying capabilities. Claude Code is therefore not a fixed model version.
A useful instruction might be: investigate why the form rejects an empty phone number, fix the bug and show which existing check confirms the behaviour. This provides a concrete goal and a testable outcome. The same approach helps with documentation and reviewing a small change.
Tools, computer use and agents
A tool gives a model access to a specific action or information source. Examples include searching, reading a file or running a command. The application receives the tool request and determines how it is executed. A model writing about a website has therefore not necessarily visited it.
With computer use, an agent can interact with software through screenshots and interface actions. A suitable environment must be connected. This does not mean every conversation automatically gains access to your computer. An AI agent uses such capabilities across multiple steps to pursue a goal.
For example, an agent gathers data from an authorised folder and prepares a report. It can check whether all files have been processed and flag missing input. Quality depends on the task boundaries, useful tools and a checked completion, as well as the model.
Which Claude should I choose?
Start with the task. Haiku suits fast, clearly bounded processing. Sonnet is a practical candidate for writing, code and everyday work with regular feedback. Opus targets more demanding work; the current API documentation recommends Opus 5.5 as a starting point for most workloads. That is vendor guidance, not a universal optimum.
Fable targets deep reasoning and tasks that must remain coherent over a long sequence. Documentation mentions it for particularly difficult work or when Opus with more reasoning effort does not meet your evaluation criteria. Mythos is not an ordinary additional subscription option: access requires verification through specific programmes.
Begin with a representative task and assess correct output, waiting time, correction work and total cost. A smaller model may handle a structured task very well; a more capable model may find useful connections in ambiguous work. Differences are situation-dependent and change between generations.
Current families compared
| Family | Orientation | Relative speed | Capability profile | Typical use | Current status |
|---|---|---|---|---|---|
| Sonnet | Everyday, iterative work | Fast | Broadly capable | Text, code, agents | 5.5 · active |
| Opus | Demanding professional work | Moderate | Complex analysis and execution | Research, code, documents | 5.5 · active |
| Fable | Deep, sustained reasoning | Slower | Difficult tasks over many steps | Extensive analysis and agents | 5.1 · active |
| Mythos | Verified programme work | Slower | Same model basis as Fable; different safeguards | Controlled specialist use | 5.1 · verification required |
| Haiku | Fast, bounded processing | Fastest | Focused language and code tasks | Classification, translation, summaries | 5.5 · active |
Access, subscriptions and model status
The assistant, API, Claude Code and cloud providers each have their own access and usage conditions. Check the current Claude pricing page and the models available in your account. This entry does not quote fixed subscription prices: a monthly amount also does not automatically describe API usage, extra credits or contract terms.
All five families in the table are active on the review date. This does not mean that every historical version still works. Lifecycle documentation distinguishes active, legacy, deprecated and retired. Deprecated means a version is being phased out; retired means requests to that version are no longer processed.
For example, Sonnet 4.5 has been deprecated since 30 September 2026, with retirement announced for 30 November 2026. Haiku 5.5 was released on 7 October 2026 and is listed as the latest active version. The older Haiku 4.5 also remains active. Consult current API documentation before copying old model names into an integration.
A useful approach for everyday users
First give Claude the relevant facts and ask it to identify missing information explicitly. Then specify an output: a comparison, draft email, table or code change. For important conclusions, request references to the original source and check that figures and exceptions are reproduced correctly.
Keep representative examples for recurring work. When switching models, compare ordinary cases and difficult exceptions. A new version may handle some tasks better and respond differently to existing instructions. A short practical check reveals whether the change actually helps your work.