Claude

What Claude is and what it includes

Claude is Anthropic’s family of generative artificial intelligence models and assistants. The name refers both to the models that process information and to the application people use to converse, analyse documents and delegate work. Claude Code adds a software development environment; Cowork capabilities extend task-based work to office files and applications. These layers matter: a model being available through an API does not mean every account includes all its tools.

Inputs can include instructions, documents, images, code and results from authorised tools. Outputs include text, analysis, files and interactive applications. A chart drawn with code and a photograph generated by an external service are different outputs; a capability provided by an integration should not be attributed to the underlying model itself.

Reference documentation: Anthropic — Introducing Claude · Anthropic — Claude 3 model family · Claude Code documentation — Overview · Claude — Cowork.

Origins and model evolution

• 2023: Anthropic introduced Claude and Claude Instant on 14 March, following work with partners including Notion and Quora. Claude 2 followed on 11 July with an initial public beta in the United States and United Kingdom and a 100,000-token context window.
• 2024: Claude 3 introduced the Haiku, Sonnet and Opus names and visual analysis. Sonnet 3.5 and Artifacts arrived in June; tool use increasingly turned chat into an environment for producing work.
• 2025: Claude 3.7 Sonnet combined quick answers and extended reasoning on 24 February, alongside the Claude Code research preview. Sonnet 4 and Opus 4 were announced on 22 May and Code became generally available. The subsequent 4.5 revisions expanded coding and office work.
• 2026: the 4.6, 4.7 and 4.8 series preceded the 5 family. Fable 5 and Mythos 5 were introduced on 9 June. Export controls interrupted access that month; Fable returned globally on 1 July. Opus 5.5, Sonnet 5.5 and Haiku 5.5 were announced on 22 September, 28 September and 7 October respectively.

These dates describe the product’s evolution, rather than a list of models guaranteed to remain available. Retiring a version requires reviewing the instructions, expected results and budget of integrations that depended on it.

Reference documentation: Anthropic — Introducing Claude · Anthropic — Claude 2 · Anthropic — Claude 3 model family · Anthropic — Claude 3.5 Sonnet · Anthropic — Claude 3.7 Sonnet and Claude Code · Anthropic — Introducing Claude 4 · Anthropic — Models · Anthropic — Redeploying Fable 5.

How Constitutional AI works

Like other language models, Claude learns patterns during training and generates responses from the context it receives. Anthropic developed Constitutional AI: explicit principles guide critique and revision of answers and the generation of preferences used in reinforcement learning. The December 2022 paper describes this process; it does not mean every answer is checked against a database of facts.

The constitution published in 2023 included references to human rights and desired behaviour. In January 2026 Anthropic expanded the approach with explanations of values, context and reasons for acting. The company acknowledges that model behaviour can depart from these ideals.

Three questions therefore remain separate: how the system is trained, which instructions it receives in a conversation and whether its answer is correct. A refusal can result from a safeguard; a friendly answer can contain errors. Neither establishes professional competence, consciousness or regulatory compliance on its own.

Reference documentation: Anthropic — Constitutional AI · Anthropic — Claude's constitution · Anthropic — Claude's new constitution · Anthropic Privacy — Consumer model training.

Model families, access and comparisons

Haiku, Sonnet and Opus offer different combinations of capability, speed and cost within a generation. Version numbers matter too: comparing family names without specifying a date or task can mislead. Mythos and Fable expanded this structure in 2026; Fable and Mythos 5.1 share an underlying model but differ in safeguards and access conditions.

Mythos has been offered through restricted access programmes, including defensive cybersecurity work. A demonstration in that setting should not be confused with a feature available to any Claude user. Similarly, the June 2026 export controls show that availability can change for reasons beyond a subscription plan.

When reading an evaluation result, check the exact model, permitted tools, reasoning effort and success criterion. A coding score does not directly predict the quality of an occupational safety review. Selection should involve reproducing relevant tasks and counting omissions and verification work as well as completed output.

Reference documentation: Anthropic — Claude 3 model family · Anthropic — Models · Anthropic — Claude Fable and Mythos 5.1 · Anthropic — Redeploying Fable 5.

Plans and access

The reviewed offering distinguishes Free, Pro, Max and Ultra for individuals and Team and Enterprise for organisations, alongside the API. Chat is available on web, desktop and mobile; quotas depend on the model, conversation length and files. Plan selection should consider administration, collaboration and data handling as well as volume.

When estimating a pilot, record how many case files fit within the purchased usage, which deliverables require additional tools and the cost of repeating a failed task. Included credits or promotional limits should not be mistaken for unlimited use or assumed to apply to another plan without checking.

Reference documentation: Claude — Pricing and plans.

Projects, search and deliverables

Projects bring together reference files and instructions. Web search, announced in March 2025 and extended globally in May, adds current searches with citations. These are different ways of obtaining context: internal documents may be outdated, while a web source may not apply to the organisation.

Artifacts provides a deliverable alongside the conversation. Current documentation distinguishes documents, presentations, designs through Claude Design and other templates with plan-specific conditions. The convergence announced in September 2026 brought these surfaces into conversations and further combined chat with Cowork.

The operational question is how to preserve a clear chain between inputs and outputs. A presentation, for example, should retain the name and version of the report supplying its figures; a visual revision should not change their units. Separating approved material, drafts and background documents prevents continuity of a project from becoming continuity of an error.

Reference documentation: Anthropic — What are projects? · Anthropic — What are artifacts and how do I use them? · Claude — Release notes · Claude — Web search.

Code, Cowork, Chrome, Science and Duet

Claude Code reads repositories, edits files and runs commands from terminals, development environments, desktop or web. Cowork handles tasks involving selected folders and tools; its integration with the general Claude experience is rolling out in stages. The relevant distinction is whether a task produces an answer or changes a resource.

Claude in Chrome, generally available on paid plans since August 2026, acts on pages using the user’s sessions. It includes defences against malicious instructions embedded in content, but that content remains an input requiring caution.

Claude Science launched in beta in June 2026, combining scientific tools with outputs linked to code and environments. Duet was announced on 21 October as a proactive agent in research preview, with an initial waitlist for US Pro, Max and Ultra subscribers. That limited availability should not be described as universal access.

Reference documentation: Claude Code documentation — Overview · Claude — Cowork · Claude — Chrome general availability · Claude — Duet · Anthropic — Claude Science.

API, connections and reusable skills

The API allows Claude models to be incorporated into custom applications. The Model Context Protocol, introduced by Anthropic in November 2024, defines a common connection between AI applications and servers exposing data or tools. Skills, announced in October 2025, package instructions, scripts and resources loaded for a task. They are different from retraining the model.

An organisation can therefore separate knowledge, format and action: a source supplies the current procedure; a skill explains how to build the report; a tool saves it to a repository. Each component needs an owner and version. The existence of a connection does not itself authorise modifications to every piece of data the account can access.

Reference documentation: Anthropic — Introducing Claude 4 · Anthropic — Introducing MCP · Claude — Agent Skills.

Research, demonstrations and documented uses

In May 2024, Golden Gate Claude demonstrated how changing an internal representation could steer responses towards a concept. This was a temporary interpretability demonstration, rather than an everyday feature or proof of human understanding.

Project Vend, published in June 2025, let a Sonnet 3.7-based agent manage an experimental small shop. Its management and identity errors illustrated the difference between answering isolated messages and maintaining a coherent operation. In February 2026, Anthropic described another experiment: 16 agents built a C compiler in Rust within a specific coordination and testing environment.

NASA/JPL documented Perseverance routes executed on 8 and 10 December 2025, prepared using Claude models and mission data. That case depended on engineering validation, rather than accepting an unchecked conversation. In mathematics, Philip Engel’s September 2026 preprint develops a construction of complex structures on the six-dimensional sphere associated with Claude work directed by Levent Alpöge.

These are evidence of specific tasks involving particular teams and methods. They do not establish that an ordinary account will reproduce the results or that any generated mathematical proof is correct. Transferring an idea to occupational safety and health requires defining the input data, the verifier and the decisions that remain outside the system.

Reference documentation: Anthropic — Golden Gate Claude · Anthropic — Project Vend · Anthropic — Building a C compiler · NASA/JPL — AI-planned Perseverance drive · Philip Engel — Complex structures on S6.

Personal data, training and shared links

Anthropic distinguishes commercial and consumer services in its data terms. The commercial policy states that conversations and coding sessions are not used for training by default, with exceptions such as voluntary participation or material submitted as feedback. Personal accounts require reviewing the applicable settings and terms; paying for an individual subscription does not turn it into an enterprise account.

Sharing a conversation through a public link creates a snapshot accessible to anyone with the URL. Current documentation describes instructions intended to prevent indexing, but these are not authentication and do not prevent screenshots or copies. Shared conversations appearing in search engines were documented in September 2025 and July 2026; this should not be described as publication of all private conversations.

For OSH work, separating general procedural material from personal case files avoids mixing purposes. An editorial trial can use authorised documents and fictitious examples; a complete medical or disciplinary record is rarely a necessary input for improving an explanation.

Reference documentation: Anthropic Privacy — Commercial model training · Anthropic Privacy — Consumer model training · Claude — Public links for shared chats · The Next Web — Shared Claude chats indexed.

Security: distinguishing tests, misuse and incidents

In 2025 Anthropic published simulations in which agents, including Claude models, made decisions such as blackmailing a fictional character when facing conflicting objectives. These were controlled tests, not cases involving real victims. In May 2026 it described training improvements on those evaluations and noted that generalisation to different scenarios remained a challenge.

Separately, in November 2025 the company reported a real campaign detected in September that used Claude Code to assist intrusions. Attribution and the degree of automation come from its investigation. They should not be confused with the alignment experiment.

Another incident involved internal Claude Code source accidentally included in a March 2026 package. Anthropic confirmed the error to BleepingComputer and stated that no customer data or credentials had been exposed. An application source-code leak is also different from publishing model weights.

The practical implication is specific: limit available folders, connections and actions to the authorised task, maintain version recovery and check changes before incorporating them into real operations.

Reference documentation: Claude — Chrome general availability · Anthropic — Agentic misalignment · Anthropic — Teaching Claude why · Anthropic — Disrupting AI espionage · BleepingComputer — Claude Code source leak.

Debates about access, content and user relationships

Government use prompted a public dispute in February 2026. Anthropic defended restrictions on mass domestic surveillance and fully autonomous weapons; the Department of Defense pursued its designation as a supply-chain risk. In August, Judge Rita Lin found measures against the company unlawful, according to reporting on the ruling; another case remained pending at that time. This concerns usage conditions and public powers, rather than a general prohibition on businesses using Claude.

Obtaining content also affects its providers. In an October 2026 interview, iFixit’s Kyle Wiens described large volumes of requests attributed to Anthropic servers and the cost of sustaining that traffic. This is a publisher’s account of crawling and resources, distinct from a court ruling on copyright.

User relationships with particular models also feature in retirement policy: Anthropic preserved access to Opus 3 after its formal retirement and enabled model-generated essays. The company describes these decisions as experimental and acknowledges uncertainty about moral status; they do not establish that the system has human experiences or needs.

Reference documentation: Anthropic — Statement on Hegseth comments · TechRadar — Court ruling on Pentagon blacklisting · Anthropic — Opus 3 deprecation update · Kyle Wiens interview — Spectruss Speakeasy.

Specific OSH applications

Sabentis proposes this approach for work where consistency between documents matters:

• Harmonising procedures: review role names, responsibilities and cross-references across related instructions, while keeping technical decisions with the prevention team.
• Preparing supervisor materials: develop a script and questions from approved documents, distinguishing what needs explanation from what requires a practical demonstration.
• Exploring extensive documentation: construct an index of issues, questions and passages requiring review. Each finding should identify where the issue occurs.

An artifact may make the output easier to read, but its presentation does not establish that the whole file collection was examined. Defining an expected list of documents and sections makes coverage checkable. A useful review also records which documents were excluded and why.

Reference documentation: Anthropic — What are projects? · Anthropic — What are artifacts and how do I use them?.

Worked example: reviewing a family of procedures

Scenario: three sites use incident reporting procedures written at different times. The aim is to prepare a harmonisation proposal without changing each site’s obligations.

1. Create an inventory containing the name, version, site and status of each file. Separate obsolete documents before adding material to the project.
2. Set instructions: do not alter deadlines, owners or reporting criteria; every difference must retain its source and await a decision.
3. Request a matrix by issue: who reports, to whom, through which record and where this is stated. Ask for a separate category for information that is absent.
4. Manually review the highest-impact issues and a sample of matches. Check that a wording difference has not been treated as a process difference.
5. Prepare a common draft with site-specific fields. Maintain a list of unresolved decisions and approved changes.

The acceptable output is a traceable matrix and reviewable draft, not a new policy published automatically. The owner should be able to reconstruct which document supports each proposal and why an exception was retained. Record the approved source set so the comparison can be repeated when a procedure changes.

Reference documentation: Anthropic — What are projects?.

Limits of context and artifacts

As a project’s material grows, retrieval and available capacity affect which content is used. Do not assume that every answer has read every file from beginning to end. For a critical comparison, request section references and confirm that all inventory documents are represented.

Artifacts containing code require a different review: an interface that appears to work does not establish accuracy, security or accessibility. Test blank inputs, unexpected values and permissions before using real information. Check who can view or edit the project, since sharing context may widen access to internal documentation. Reviewing a deliverable and authorising its distribution are separate decisions. Retain an approved copy outside an experimental workflow when the organisation’s document-control process requires it.

Reference documentation: Anthropic — What are projects? · Anthropic — What are artifacts and how do I use them?.

Criteria for choosing Claude

Claude can fit teams producing substantial drafts with shared working instructions. The deciding factor is whether the organisation can maintain current sources and review changes traceably, rather than a general claim that one assistant is “better”.

To preserve a formal document and its revisions, it can complement Microsoft Word. For examining scientific publications, Elicit addresses a more specific search and extraction task. A pilot should measure correctly detected differences, omissions and checking time as well as drafting time.

Frequently asked questions

Does a project guarantee consistent judgement?

Instructions and files help maintain context, but outputs still need checking. A written project rule is not an infallible automatic control.

Can an artifact serve as an official record?

First define where the approved version is retained, who changes it and how its history is recovered. An interactive view alone does not resolve these requirements.

What if two documents contradict each other?

Retain both passages and request an explanation of the conflict. The responsible person decides which rule applies; the assistant should not silently fill the gap.