{
  "id": "claude",
  "title": "Claude",
  "lang": "en",
  "source_lang": "en",
  "summary": "Claude is Anthropic’s family of generative AI models and assistants for language, visual analysis, coding, and tool-assisted tasks.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "product",
  "aliases": [
    "Claude AI",
    "Anthropic Claude"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/claude/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/claude/",
      "json": "https://aiwiki.page/zh/claude.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Claude is a family of generative artificial intelligence models and associated assistant products developed by Anthropic. Introduced publicly in March 2023, it supports activities such as answering questions, analyzing documents, writing and revising text, generating software, and interpreting images. The name refers both to the underlying models, which developers can access through software interfaces, and to applications that present those models as conversational assistants.\n\nClaude belongs to the class of large language models: systems trained to learn statistical relationships in large collections of data and generate outputs conditioned on instructions and context. Its development is closely associated with Constitutional AI, Anthropic’s approach to using explicit principles and model-generated feedback in the training of assistant behavior. Later versions extended the original text-based service with visual input, longer working contexts, tool use, and interfaces for software development.\n\nUnderstanding Claude requires distinguishing the model from the surrounding product. A model generates responses; a deployed service also manages files, conversation history, permissions, external tools, and user accounts. Capabilities therefore depend on the particular model version, interface, and configuration rather than on the Claude name alone."
    },
    {
      "heading": "Origins and development of the model family",
      "text": "Anthropic was founded in 2021 by a group that included former OpenAI researchers Dario Amodei and Daniela Amodei. The company’s research program connected the development of capable language models with questions about their reliability, interpretability, and control. Claude became the principal commercial expression of that program, combining a conversational interface with access for organizations building their own applications.\n\nClaude’s public introduction on 14 March 2023 included the original Claude model and Claude Instant, a lower-latency offering intended for comparatively economical use. The initial service focused on text generation, dialogue, summarization, and related natural language processing tasks. In May 2023, Anthropic announced a context capacity of approximately 100,000 tokens, making large document collections a prominent part of Claude’s early positioning. Claude 2 followed on 11 July 2023, alongside a public-facing web experience initially offered in the United States and United Kingdom. Claude 2.1, announced on 21 November 2023, increased the advertised context capacity to 200,000 tokens and introduced a beta tool-use capability.\n\nThe Claude 3 family, announced on 4 March 2024, established three model names: Haiku, Sonnet, and Opus. These designated different intended balances among speed, cost, and capability rather than different subject areas. Claude 3 also added image understanding to the family’s principal capabilities. Claude 3.5 Sonnet, introduced on 20 June 2024, was presented as improving performance across several evaluations while retaining the Sonnet product tier. An updated Claude 3.5 Sonnet released in October 2024 accompanied a public beta of computer-use functionality.\n\nClaude 3.7 Sonnet, announced on 24 February 2025, introduced an optional extended-thinking mode. Anthropic described it as a hybrid reasoning model because the same model could provide comparatively immediate answers or allocate additional computation to a problem before producing its final response. The announcement also introduced Claude Code as a research preview for terminal-based software-development work.\n\nThese releases illustrate why Claude is better treated as an evolving model family than as a single fixed system. A statement about one generation’s context limit, image support, or tool behavior does not necessarily apply to another. Model names, dated identifiers, and release-specific documentation provide more precise descriptions than the family name. Product availability is a separate matter: an announced model may be accessible through one service before another, and a previously available version may later be retired."
    },
    {
      "heading": "Language modeling, training, and generation",
      "text": "Claude’s basic operation follows the language-modeling approach used by contemporary generative assistants. Input text is converted through tokenization into numerical units representing words, word fragments, punctuation, or other symbols. A trained artificial neural network processes these units and assigns probabilities to possible continuations. Claude’s internal architecture, model sizes, and complete training mixtures have not been publicly specified for every release, so descriptions of the family should not assume a disclosed parameter count or a single unchanging implementation.\n\nFor a sequence of tokens \\(x_1,\\ldots,x_T\\), autoregressive language modeling represents its probability as\n\n\\[P(x_1,\\ldots,x_T)=\\prod_{t=1}^{T}P(x_t\\mid x_1,\\ldots,x_{t-1}).\\]\n\nAt generation time, the system repeatedly selects a continuation token and incorporates it into the context for subsequent predictions. This mechanism can produce an explanation, a program, a translation, or a structured tool request because all can be represented as sequences. Decoding settings influence how alternatives are selected, but reducing randomness does not make a response necessarily correct.\n\nThe training process distinguishes broad pretraining from later behavioral adaptation. Pretraining uses large quantities of training data to learn linguistic patterns, factual associations, programming conventions, and other regularities. Anthropic has described its training sources in broad categories that include publicly available information, licensed or otherwise supplied data, and internally generated data. Such descriptions do not constitute a complete inventory of individual documents. Subsequent fine-tuning and feedback-based training shape how the resulting system follows instructions, handles uncertainty, and responds to requests.\n\nConversation is an application of this sequence-processing machinery. Instructions, user messages, previous assistant responses, and sometimes tool results are assembled into a structured input. The model generates a response conditioned on that assembled material. Apparent continuity across turns therefore depends substantially on what the application supplies again, not on a persistent human-like awareness of earlier interaction.\n\nA useful distinction is between information encoded in model parameters and information supplied at runtime. Learned parameters may support an answer about a historical event without an external lookup. By contrast, a recently uploaded report can be analyzed because its contents are present in the input, even if it was created after training. Neither route guarantees accuracy: parameter-based answers can contain outdated associations, while context-based answers can misread or overlook evidence.\n\nDeployment also involves neural-network inference infrastructure, including accelerators, memory management, and scheduling. These systems affect throughput, latency, and operating cost. Consequently, improvements in a Claude service can arise from changes to the underlying model, the inference system, or the application layer, and these changes need not occur together."
    },
    {
      "heading": "Constitutional AI and behavioral alignment",
      "text": "Claude is closely associated with Anthropic’s research on AI alignment, particularly Constitutional AI. Introduced in a 2022 research paper, the method uses a written set of principles to guide the production and evaluation of model responses. The term “constitution” describes those guiding principles; it does not indicate that the model has legal authority or that its behavior is determined by an infallible rule engine.\n\nThe original research described a supervised stage in which a model generated responses to prompts, criticized those responses against selected principles, and produced revisions. The revised examples were then used for further training. A principle might concern avoiding harmful assistance, respecting privacy, or responding without unnecessary judgment. The important mechanism is that a model helps create the behavioral examples from which another training stage learns, reducing dependence on individually written human corrections for every example.\n\nA second stage used AI-generated preferences. A model compared candidate responses with reference to constitutional principles, and those comparisons helped train a preference model. Reinforcement learning then optimized the assistant toward responses favored by that preference model. This approach is often called reinforcement learning from AI feedback. It is related to reinforcement learning from human feedback, but differs in how a substantial part of the comparative feedback is produced. Human decisions remain involved in selecting principles, designing experiments, constructing evaluations, and determining deployment policy.\n\nConstitutional training does not mean that every production answer is generated by visibly consulting a fixed checklist. Its influence can be incorporated into the model’s learned behavior through training. Runtime instructions and safety systems may provide additional controls, but they are distinct from the training procedure. Nor does the publication of a Constitutional AI method establish that every later Claude model uses exactly the same recipe, datasets, or constitution.\n\nThe approach also creates substantive design questions. Broad principles can conflict: answering helpfully may compete with withholding dangerous operational detail, while avoiding unjustified certainty may make an answer less direct. Principles stated in natural language require interpretation, and models can interpret them inconsistently. Agreement with a preference model is therefore not equivalent to proof of factual correctness or ethical adequacy.\n\nClaude’s refusals, qualifications, and conversational style emerge from this wider training and deployment process. A refusal can prevent harmful output, but an excessively broad refusal can also block a legitimate request. Conversely, a courteous and cautious tone may accompany an incorrect answer. Behavioral alignment and factual reliability overlap, but they are not interchangeable properties and require different kinds of evaluation."
    },
    {
      "heading": "Context windows, document analysis, and retrieval",
      "text": "A context window is the bounded amount of input and generated material a model can process within a particular request. Claude became notable early in its development for large context capacities: approximately 100,000 tokens were announced in May 2023, and Claude 2.1 subsequently supported 200,000. These are token counts, not fixed word or page counts. The number of tokens required for a document varies with language, formatting, code content, and the tokenizer.\n\nLong context supports tasks such as comparing contracts, summarizing research collections, tracing references across technical documentation, or examining a substantial codebase. Instead of reducing every source to a short summary before asking a question, an application can supply large portions of the original material. This can preserve qualifications and cross-references that might otherwise disappear during preprocessing. However, the nominal capacity is an upper limit on supplied material, not a guarantee that every passage will receive equally effective attention.\n\nA model may fail to retrieve an important detail from a long input, confuse similarly named entities, or merge claims from separate documents. Position, repetition, conflicting evidence, and the wording of the question can all affect results. Locating one distinctive sentence in a long document is also a different task from integrating evidence scattered across many sections. Tests of long-context performance therefore need to distinguish simple retrieval, multi-document comparison, and reasoning that depends on several linked facts.\n\nAn alternative or complementary technique is retrieval-augmented generation. In this arrangement, a separate information retrieval system selects passages from a collection and supplies them to the model. A database or search index can hold far more information than fits into one context window. Retrieval narrows the input to material judged relevant, whereas long context allows more of that material to be considered together. Neither approach eliminates error: retrieval can omit a necessary source, and generation can misrepresent a correctly retrieved passage.\n\nDocument handling also depends on preprocessing. A PDF may contain selectable text, scanned pages, diagrams, tables, or a mixture of all four. Its visible page order may not match the order extracted by software. Images can preserve spatial relationships that plain text loses, while text extraction may capture small print more efficiently. The Claude model sees the representation passed to it, not an abstract, perfectly reconstructed document.\n\nApplication-managed memory is another separate mechanism. A service may retain instructions, project files, or selected conversation information and insert them into later requests. Such features should not be confused with unlimited model context or continuous retraining on an individual user’s activity. Persistent storage, retrieval, and context construction are product functions whose behavior can change independently of the model."
    },
    {
      "heading": "Multimodal input and generated workspaces",
      "text": "The Claude 3 family introduced image input alongside text, extending Claude into multimodal applications. Supported tasks include describing a photograph, extracting information from a chart, interpreting a diagram, and answering questions about a screenshot. The output in these uses is generally text. The ability to interpret images does not, by itself, imply that the same model can generate original images or perform every kind of audio and video processing.\n\nVisual analysis requires more than naming objects. A chart question can depend on recognizing axes, reading labels, distinguishing line styles, and relating quantities across the figure. A screenshot can combine icons, text, layout, and application state. A model may correctly recognize the overall subject while making a numerical or spatial error. Small text, low resolution, dense tables, and ambiguous visual encodings can increase the difficulty. Performance is consequently sensitive to the actual input representation, not merely to the subject of the image.\n\nClaude’s visual capability also supports reasoning across images and prose. A user can provide a diagram together with a written specification and ask whether the two agree. In software work, an interface screenshot can serve as a reference for generating front-end code. In document analysis, an embedded figure can be interpreted alongside its caption. These tasks combine visual interpretation with language generation; they do not provide a guarantee that geometrical measurements, extracted values, or inferred relationships are exact.\n\nIn June 2024, Anthropic introduced Artifacts, an interface feature that separates certain generated outputs from the conversational stream. Code, formatted documents, diagrams, and interactive web content can appear in a dedicated workspace where successive revisions remain easier to inspect. This changes the interaction pattern from receiving a sequence of isolated answers to iteratively developing an object. A request to revise a chart or modify an interface can refer to an existing artifact rather than requiring the user to reconstruct it manually.\n\nArtifacts illustrate the distinction between content generation and execution. A model may produce HTML, styling, or JavaScript, while the application supplies the preview environment. Whether the content runs correctly depends on browser behavior, dependencies, permitted interfaces, and execution restrictions. A plausible-looking preview is not evidence that the underlying program satisfies all requirements or is safe for unrestricted deployment.\n\nThese interfaces also raise questions in human–computer interaction. Separating an output from the chat can make revision history and application state more legible, but generated interfaces may conceal errors behind a polished presentation. Accessibility, keyboard behavior, and error handling are properties of the resulting artifact, not automatic consequences of using a visually capable model."
    },
    {
      "heading": "Tool use, computer interaction, and coding agents",
      "text": "Tool use allows Claude to request actions outside ordinary text generation. In an application programming interface workflow, a developer supplies descriptions of available tools and their expected arguments. The model can produce a structured request, the host application executes the corresponding operation, and the result is returned as additional context. A tool might query a database, perform arithmetic, retrieve a document, or interact with another service.\n\nThe model’s request is not itself the external action. This separation matters because the application controls which tools exist, what credentials they use, and whether proposed operations require approval. A weather query and a command that deletes files may have a similar conversational form while carrying very different consequences. Tool schemas constrain formatting, but they do not establish that the selected action is appropriate or that the arguments reflect the user’s intent.\n\nIn October 2024, Anthropic introduced computer use in public beta with an updated Claude 3.5 Sonnet. The demonstrated interaction cycle involved receiving screenshots and requesting actions such as moving a pointer, clicking, or typing. The model could then inspect a subsequent screenshot and continue. This extends tool use to software without a purpose-built semantic interface, but it also introduces fragility: screen coordinates, delayed loading, unexpected dialogs, and changes in layout can disrupt an action sequence.\n\nAn AI agent built around Claude typically combines this repeated observation-and-action loop with a task objective, working state, and termination conditions. Reliability becomes a property of the entire loop. Even if each individual action is usually correct, a long chain can accumulate failures. Robust implementations distinguish reversible inspection from consequential modification and maintain records of actions and tool results.\n\nClaude Code, introduced as a research preview in February 2025, applied this pattern to software development from the terminal. It could inspect project files, propose or make edits, run commands, and use development tools within the permissions of its environment. Such a coding agent differs from a chatbot that merely returns a code block: it can work with the actual repository and observe the consequences of a change.\n\nThe resulting workflow connects generation with version control, testing, and code review. A successful edit requires more than syntactically valid code; it must fit the project’s dependencies, conventions, and intended behavior. Passing a test suite demonstrates compatibility with the tested cases, not complete correctness. Tool-assisted coding therefore exposes both the practical value of language models and the limits of treating generated patches as self-validating results."
    },
    {
      "heading": "Applications, deployment, and commercial access",
      "text": "Claude reaches users through several distinct delivery channels. The consumer-facing assistant provides a conversational interface, file handling, and product features such as projects or generated workspaces. Developer access exposes model capabilities through an API. Enterprise arrangements add organizational administration and integration requirements. These channels share underlying models in some configurations, but they do not necessarily offer identical features, limits, or data-handling terms.\n\nAnthropic has also distributed Claude through major cloud computing platforms, including Amazon Bedrock and Google Cloud’s Vertex AI. Such arrangements allow organizations to incorporate Claude into an existing cloud environment rather than build directly against only Anthropic’s own service. The hosting route affects authentication, billing, available regions, and integration with surrounding infrastructure. It does not imply that a model has unrestricted access to other data held in the same cloud account.\n\nThe Haiku, Sonnet, and Opus naming scheme provides a commercial way to distinguish models by intended performance and resource trade-offs. A fast, lower-cost model may suit high-volume classification or short extraction tasks, while a more capable model may be chosen for difficult analysis. These labels do not establish an immutable ranking across every generation. A newer model in one tier can outperform an older model in another on particular tasks, and latency depends on request length, load, and deployment configuration.\n\nAPI economics are commonly expressed in terms of input and output tokens. A long source document increases input processing, while an expansive answer increases output generation. Long-running agent workflows can incur additional usage through repeated requests and tool-result processing. Features such as prompt caching and batch processing can change the cost structure, but their availability and exact behavior are service-specific. Operational evaluation therefore considers not only answer quality but also response time, concurrency, and total task cost.\n\nCommon applications include customer-support assistance, document extraction, research support, internal knowledge search, writing, and programming. In each case, the surrounding software defines the actual workflow. A support system may retrieve approved policy text before asking Claude to draft a response; a document pipeline may validate extracted fields against a schema; a coding service may run tests after applying a patch. These additions can substantially change practical reliability.\n\nClaude is proprietary software, and its model weights have not generally been distributed as open-source software. Organizations using the service consequently depend on external model availability, version policies, and contractual arrangements. A fixed model identifier can help stabilize an application, but reproducibility also requires recording prompts, tool definitions, decoding settings, and relevant product behavior."
    },
    {
      "heading": "Evaluation, reasoning, and factual reliability",
      "text": "Claude’s capabilities are evaluated through a mixture of standardized benchmarks, task-specific tests, human judgments, and real-world use. Benchmark categories include general knowledge, mathematical problem solving, programming, visual understanding, instruction following, and long-context retrieval. A reported score describes performance under a particular evaluation procedure; it does not establish a universal level of intelligence or guarantee the same success rate in an application.\n\nEvaluation details can substantially affect results. A model may receive one attempt or several, use tools or answer unaided, and operate with short or extended reasoning budgets. Software benchmarks may differ in whether dependency installation, repository inspection, or repeated test execution is permitted. Comparisons are most meaningful when these conditions are aligned. A model release announcement that combines scores obtained under different conditions cannot be interpreted as a single controlled experiment.\n\nThe extended-thinking mode introduced with Claude 3.7 Sonnet makes computational budget an explicit part of this picture. Allocating more computation before the final answer can improve some difficult tasks, but it generally increases latency and resource use. The resulting explanation is not a certified transcript of every internal process, and a longer reasoning response is not evidence of correctness. A mathematical solution still depends on valid steps, and a software fix still depends on the behavior of the edited program.\n\nLike other language models, Claude can produce an AI hallucination: a fluent claim unsupported by the supplied evidence or by the facts. Examples include invented citations, nonexistent software functions, and mistaken summaries. The underlying generation process rewards plausible continuations learned during training; it does not automatically perform an external verification for each sentence. Retrieval and tool use can supply evidence, but the model can still misinterpret the result or attribute a statement to the wrong source.\n\nBenchmark validity also depends on separating training material from evaluation material. Benchmark data contamination can inflate performance when questions, solutions, or close variants appear in training. Even without direct overlap, repeated development against a familiar benchmark may reduce its value as a measure of unfamiliar tasks. Private test sets and newly constructed evaluations address parts of this problem, although they create their own issues of transparency and repeatability.\n\nFor deployed applications, task-level measures often matter more than broad rankings. An extraction system can be assessed for missing fields and false positives; a coding agent for resolved issues and introduced regressions; a research assistant for accurate attribution. Probability calibration is another distinct concern: confident wording is not a calibrated numerical estimate of correctness, and an assistant’s willingness to answer does not independently establish that its answer is reliable."
    },
    {
      "heading": "Security, privacy, and system-level constraints",
      "text": "Claude’s integration with documents, external tools, and organizational systems creates security questions beyond the content of a single answer. A major example is prompt injection, in which untrusted material contains instructions intended to redirect the assistant. A retrieved web page might tell the model to ignore its task, disclose confidential information, or invoke an unrelated tool. The attack exploits the fact that both legitimate instructions and hostile text can enter the model as language.\n\nThe distinction between instructions and data is therefore a security boundary, not merely a formatting convention. A document may need to be summarized without being allowed to redefine the assistant’s permissions. Systems commonly label message roles and tool outputs, restrict available actions, and validate requests outside the model. Nevertheless, a natural-language boundary is not equivalent to a formally enforced isolation mechanism. Model-level resistance can reduce risk without eliminating the need for application controls.\n\nTool permissions determine the consequences of a successful manipulation. An assistant limited to reading public pages has a different risk profile from one able to send messages, alter production records, or execute shell commands. Least privilege limits each component to the authority necessary for its task. Sandboxing can constrain execution, while explicit approval steps can separate a proposed action from its authorization. These controls operate in the host system rather than depending entirely on the model’s judgment.\n\nPrivacy questions also require distinguishing service configurations. Data submitted through a consumer application, a commercial API, or a cloud-hosted integration may be governed by different settings and contractual terms. Retention, model-improvement use, administrative access, and deletion are separate issues. The general statement that Claude is “private” or “uses user data” is insufficient without identifying the relevant product and policy. Information placed in a prompt must be processed by the service even when it is not used for training.\n\nThere are also limits to what model refusals can establish. A refusal is evidence about the response to a particular request, not proof that all alternative prompts or tool-mediated routes are blocked. Conversely, filtering a request can prevent benign work if context is misunderstood. Security evaluation therefore considers both successful attacks and unnecessary restrictions, alongside conventional measures such as credential protection, access logs, and incident handling.\n\nClaude-based agents also inherit the risks of their surrounding software. A correctly formatted tool request may still exploit excessive permissions, and a secure model response cannot repair a vulnerable integration. The relevant threat model includes users, external documents, connected services, execution environments, and administrators, with different trust assumptions for each. System-level assurance depends on how these boundaries are implemented and tested, not solely on the behavior of the underlying language model."
    }
  ],
  "links": [
    "generative-artificial-intelligence",
    "large-language-model",
    "constitutional-ai",
    "natural-language-processing",
    "tokenization",
    "artificial-neural-network",
    "training-data",
    "fine-tuning",
    "neural-network-inference",
    "ai-alignment",
    "reinforcement-learning",
    "reinforcement-learning-from-human-feedback",
    "context-window",
    "retrieval-augmented-generation",
    "information-retrieval",
    "database",
    "multimodal-learning",
    "html",
    "javascript",
    "human-computer-interaction",
    "application-programming-interface",
    "ai-agent",
    "claude-code",
    "version-control",
    "code-review",
    "cloud-computing",
    "open-source-software",
    "ai-hallucination",
    "benchmark-data-contamination",
    "probability-calibration",
    "prompt-injection",
    "least-privilege",
    "sandbox-computer-security"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}