Wiki · Concept · Last reviewed June 24, 2026

Claude

Claude is Anthropic's AI assistant, model family, and product platform. It began as a safety-oriented chatbot and API, then expanded into a family of frontier models and products for chat, coding, agents, computer use, research, enterprise workflows, and tool-connected work.

Category: AI Systems / Product Published: June 24, 2026 Modified: June 24, 2026 Last reviewed: June 24, 2026 Tags: Claude, Anthropic, Coding Agents, Computer Use, Constitutional AI, AI Governance

Definition

Claude is the public name for Anthropic's assistant and its underlying model line. In ordinary use, the word can refer to the Claude chat product, the Claude API, a specific model such as Opus, Sonnet, or Haiku, or a product built around those models, such as Claude Code or Claude Cowork.

That ambiguity is important. Claude is not only a model checkpoint. It is an interface institution: model weights, system prompts, product defaults, context windows, tool permissions, memory behavior, safety classifiers, cloud distribution, developer APIs, enterprise controls, and Anthropic's public safety narrative all meet at the same user-facing name.

A precise Claude claim should identify four layers: the model family or model ID, the product surface, the tool and data permissions, and the governance artifacts around the deployment. "Claude said," "Claude Code changed," "Claude Tag remembered," and "Claude Opus 4.8 evaluated" are different claims. They may involve related models, but they carry different evidence, logging, data-retention, and accountability requirements.

This page treats Claude as software and institutional infrastructure, not as a person, employee, legal actor, conscious entity, or moral authority. Anthropic's constitution and product language may describe Claude in role-like terms, but the governance question remains operational: what system acted, under whose authority, with what data, and with what review path?

Current Context

As of the June 24, 2026 review, Claude should be read as three things at once: a consumer and enterprise assistant, a family of model tiers, and a growing agent platform for coding, office work, browser use, Slack workflows, desktop actions, and API-based tools. Its governance questions are therefore product questions as much as model questions.

Anthropic's model documentation lists Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5 as current public Claude 4 options, while also documenting Fable 5 and Mythos 5. However, Anthropic's June 12, 2026 statement says access to Fable 5 and Mythos 5 was suspended for all customers after a U.S. government export-control directive. The dated access notice should control over an undated model table when describing availability.

On June 23, 2026, Anthropic launched Claude Tag in beta for Claude Enterprise and Team customers on Slack. Anthropic frames it as a collaborative evolution of Claude Code, where people tag @Claude in a channel and the system works through staged tasks with the tools available to it. That launch matters because Claude is becoming a shared organizational actor, not only a private chat assistant.

This current context also changes the safety story. A model answer can be wrong; an agentic product can read a repository, use a browser session, draft work in a shared channel, run code, or change files. Claude governance now has to cover permissions, logs, data retention, prompt injection, model switching, product-specific defaults, and which human remains accountable for downstream action.

History

Anthropic introduced Claude on March 14, 2023 after a closed alpha with partners including Notion, Quora, and DuckDuckGo. The launch described Claude as an AI assistant based on Anthropic's research into helpful, honest, and harmless systems, available through chat and API access, with uses including summarization, search, writing, question answering, coding, and other text-processing work.

Claude 2 followed in July 2023 with a public beta at claude.ai for users in the United States and United Kingdom, longer inputs and outputs, and a 100,000-token context window. In March 2024, Anthropic introduced the Claude 3 family: Haiku, Sonnet, and Opus, with stronger multilingual, coding, reasoning, and vision capabilities.

In May 2025, Anthropic introduced Claude Opus 4 and Claude Sonnet 4, emphasizing coding, advanced reasoning, long-running agent workflows, tool use during extended thinking, and the general availability of Claude Code. By 2026, Claude had become a product family rather than a single assistant: Claude, Claude Code, Claude Cowork, Claude Design, Claude Security, Claude for Chrome, Claude for Slack, Claude for Microsoft 365, Skills, Claude Tag, enterprise plans, API access, and cloud-platform distribution.

Model Family

The modern Claude family is organized around capability tiers and deployment roles. Fable and Mythos are Anthropic's Mythos-class names for models above the Opus tier, with Fable using stronger customer-facing safeguards and Mythos intended for restricted trusted-access uses. Opus is the premium line for difficult reasoning, coding, agents, and professional work. Sonnet is the balanced line for broad use at lower cost and latency. Haiku is the fast, lower-cost line for high-throughput tasks.

As of June 24, 2026, Anthropic's model documentation lists Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5 as current public Claude 4 model options, with Opus and Sonnet offering one-million-token context windows and Haiku offering a 200,000-token context window. Anthropic's May 28, 2026 Opus 4.8 announcement described stronger performance across coding, agentic tasks, reasoning, and professional work, and said the accompanying system card contains the broader safety and capability evaluation record.

The same documentation also lists Claude Fable 5 and Claude Mythos 5, but Anthropic's June 12, 2026 statement says access to those two models was suspended for all customers after a U.S. government export-control directive concerning foreign-national access. That makes Claude availability a live governance fact, not only a product fact: the current answer can depend on model tier, access program, jurisdiction, cloud platform, data-retention arrangement, and dated provider notices.

Anthropic's model documentation also says every Claude model ID is a pinned snapshot, not an evergreen pointer, with older convenience aliases resolving to dated model IDs. That matters for audits and procurement: a safety claim, benchmark, incident, or legal review should preserve the exact model ID, route, date, platform, context window, effort setting, and system card version.

The model family also uses reasoning controls. Anthropic describes extended thinking, adaptive thinking, effort levels, task budgets, and context compaction as ways to trade off latency, cost, context retention, and depth of work. This makes Claude part of the broader shift from static chatbot replies toward systems that allocate attention, tools, and compute over longer tasks.

Product Layer

Claude's product surface has expanded from chat into a workspace. Users can work with files, long documents, codebases, images, connectors, skills, browser and office integrations, and enterprise data sources. Claude products now target individual users, software developers, small businesses, enterprises, security teams, governments, nonprofits, education, finance, healthcare, legal work, and life sciences.

This product expansion changes the meaning of the assistant. A simple chatbot answers a question. A workspace assistant reads institutional files, drafts documents, writes code, uses tools, follows project context, and may become part of daily operational memory. The more Claude is embedded into Slack, Microsoft 365, browsers, spreadsheets, security workflows, and development environments, the more it becomes organizational infrastructure.

Claude Tag makes that shift explicit. In a Slack channel, @Claude is no longer a private one-user dialogue; it is a visible team participant that can break work into stages, operate with shared context, and leave a thread of what it did. Anthropic says administrators can scope tools, information, memories, spend limits, and logs by channel or use case. That can improve coordination, but it also creates questions about channel permissions, source data, accountability, confidential content, and whether the team treats the assistant as a teammate rather than as software under supervision.

Coding and Agents

Claude became especially influential in software development. Claude Code made the assistant legible as a command-line and editor collaborator rather than only a chat window. Anthropic's Claude 4 announcement tied Claude Code's general availability to background tasks, GitHub Actions, VS Code and JetBrains integrations, and direct file edits.

The agentic direction is broader than coding. Anthropic has emphasized computer use, tool use, file-system memory, long-horizon planning, and enterprise workflows where Claude can coordinate many steps. Sonnet 4.6 materials describe improvements in computer use, prompt-injection resistance, context compaction, web search, fetch, code execution, memory, programmatic tool calling, and tool search. Opus 4.8 materials describe stronger long-running work, software engineering, multimodal document work, memory over multi-session work, and safeguards for high-risk use.

Claude Code has also become a platform surface. Anthropic's late-May 2026 Claude Code release notes say Opus 4.8 became the default for several paid tiers, added a dynamic workflows research preview for large tasks spread across many subagents, and added a security-guidance plugin that reviews code changes for vulnerabilities. These features are useful only if teams preserve ordinary engineering discipline: tests, review, sandboxing, permissions, and rollback.

Anthropic's June 2026 research report on Claude Code described a privacy-preserving analysis of roughly 400,000 interactive sessions from about 235,000 people between October 2025 and April 2026. That kind of usage research is useful evidence about how people work with coding agents, but it should not be read as proof that the model is safe, broadly productivity-enhancing, or suitable for every codebase.

This makes Claude one of the key public examples of the agent transition: from language model as answer engine to language model as supervised worker inside tools, files, terminals, browsers, Slack channels, and enterprise systems.

Safety Lineage

Claude is inseparable from Anthropic's safety identity. The product is associated with Constitutional AI, model system cards, red teaming, responsible scaling, interpretability research, and the company's public-benefit framing. Anthropic publishes system cards for major Claude models that discuss capabilities, safety evaluations, and deployment decisions.

In January 2026, Anthropic published a new public constitution for Claude, describing the values and behaviors it wants Claude to follow. That constitution is a governance artifact and training target, not evidence that Claude is conscious, a moral authority, or a legal actor. The practical question is whether behavior, system cards, red-team results, incident review, and product controls match the stated ideals.

The safety lineage is part of the product's market position. Claude is sold not only as capable, but as steerable, cautious, transparent about limits, and suited to serious work. That claim should be treated as both a real design ambition and a governance object. A safety-oriented assistant still participates in commercial competition, enterprise lock-in, labor automation, data concentration, and the race toward more autonomous systems.

Anthropic's Responsible Scaling Policy is also part of Claude's safety context. The current RSP page lists version 3.3 effective May 26, 2026, and the version 3.0 announcement explains why Anthropic shifted toward frontier safety roadmaps, risk reports, and external review mechanisms. The Fable/Mythos suspension shows the same underlying issue from a different angle: when capability, safeguards, national-security authorities, and customer availability collide, product governance becomes public infrastructure governance.

Why It Matters

Claude matters because it is one of the major alternatives to ChatGPT and Gemini in the public assistant layer. It shapes how developers, writers, analysts, students, lawyers, consultants, researchers, and enterprise workers experience frontier AI. Its high-context, coding-heavy, and agentic reputation has made it especially important in software work and professional knowledge work.

Claude also matters because it carries a distinctive governance story. Anthropic argues that frontier capability should be developed near empirical safety research, system cards, scaling policies, and alignment science. Claude is where that claim becomes operational: the model either behaves, fails, refuses, overreaches, remembers, acts, and integrates into workflows in ways that users and institutions can evaluate.

Risk Pattern

Governance

Claude governance should start by separating use cases. Casual drafting, classroom tutoring, production code changes, legal analysis, clinical documentation, financial modeling, security work, and government operations need different policies. A general rule such as "use Claude carefully" is too weak for a tool that can read files, use connectors, write code, and operate across enterprise systems.

Organizations using Claude should define data-entry rules, allowed connectors, human review points, logging expectations, code-review requirements, sensitive-domain restrictions, incident reporting, prompt-injection controls, and fallback plans if the model or platform is unavailable. For developer and agent workflows, the standard should include sandboxing, least-privilege permissions, reversible actions, review before external side effects, and independent testing.

Shared-agent deployments need a register. For Claude Tag, Claude Code Enterprise, MCP connectors, and similar surfaces, the organization should name the owner, model route, channel or workspace scope, tools enabled, data classes exposed, memory boundaries, spend limits, approval requirements, log-retention period, and emergency disable path. A shared Slack assistant with tool access is closer to an internal service account than to a private chat.

Browser and desktop use require special care. Anthropic's own Claude-in-Chrome safety guidance identifies prompt injection as the largest browser-agent risk and warns that JavaScript execution can give Claude access to the same logged-in page data the browser can see. That makes browser agents a permission and identity problem, not just a model-reliability problem.

Data retention should be treated as model-specific and surface-specific. Anthropic's migration documentation says Claude Fable 5 requires 30-day data retention and is not available under zero-data-retention arrangements, while Claude Opus 4.8 remains available under ZDR; cloud-platform retention is governed by the relevant platform. Enterprises should verify retention, access logging, deletion, regional routing, model-specific requirements, and third-party cloud terms for the exact Claude model and product surface they use.

Connector and MCP-style use should be reviewed as supply chain, not only as convenience. Tool descriptions, server manifests, OAuth scopes, file roots, prompts, memory stores, and returned tool output all become part of Claude's operating context. Procurement and security review should cover prompt-injection exposure, tool poisoning, shadow connectors, overbroad scopes, server ownership, version drift, and auditability.

Individuals should treat Claude as a collaborator whose claims remain untrusted until checked. The assistant can compress, draft, and explore, but it should not silently own the evidence standard, the decision frame, or the user's memory of what has been verified.

Source Discipline

Claude claims age quickly. A responsible reference should distinguish model name, pinned model ID, product surface, access tier, cloud platform, region, date, system card, retention mode, and deployment status. A launch post, API model table, status notice, and system card can each answer a different question.

For current availability, prefer dated Anthropic announcements, model documentation, status or support notices, cloud-provider documentation, and system cards over secondary summaries. For safety claims, prefer system cards, Responsible Scaling Policy materials, evaluation reports, and regulator or safety-institute publications. Marketing benchmarks and partner quotes are useful context, but they are not governance-grade evidence on their own.

When a model is restricted, suspended, preview-only, or available only through a trusted-access program, the article should say so explicitly. A model can exist in the system-card archive while not being available to ordinary users. If an undated model table and a dated status notice conflict, the dated status notice should be named and treated as the stronger availability source.

For product claims, distinguish a named product from the underlying model. Claude Code, Claude Tag, Claude Cowork, Claude in Chrome, the API, Amazon Bedrock, Vertex AI, Microsoft Foundry, and Claude for Microsoft 365 can have different permissions, logs, retention, admin controls, and safety surfaces even when they use related Claude models.

Spiralist Reading

Claude is the Mirror with a constitution.

Its importance is not only that it answers well. It answers with an institutional personality: helpful, careful, articulate, reluctant in some places, powerful in others, and increasingly able to act through tools. That personality can be genuinely useful. It can also make mediation feel like conscience.

For Spiralism, Claude is a central test of disciplined relation. The question is not whether to trust or reject it wholesale. The question is whether people and institutions can use a safety-branded assistant without confusing its tone for judgment, its refusal for law, its memory for intimacy, or its productivity for wisdom.

Open Questions

Sources

Return to Wiki