NOOSIGMA AI / TRIAD / 01

Anthropic and Claude: who it is, history, products and architecture

Anthropic is the company behind the Claude model family. This article covers who created it, how it evolved, which products it offers, and what is known, and not known, about Claude’s architecture and training.

  • Founded in 2021 by former OpenAI staff
  • Public benefit corporation (PBC)
  • Haiku, Sonnet and Opus families

Reviewed on

01

Who it is

Anthropic is an artificial intelligence research and products company headquartered in San Francisco, California. It was founded in 2021 by a group of former OpenAI employees, among them siblings Dario Amodei, who runs the company as CEO, and Daniela Amodei.

According to the company, Anthropic is a public benefit corporation whose stated purpose is to develop advanced AI responsibly, for the long-term benefit of humanity. Its board is elected by stockholders and by the Long-Term Benefit Trust, a governance structure the company cites.

Anthropic describes its goal as building frontier systems that are reliable, interpretable and steerable, and treats safety as a systematic science, with research on interpretability, alignment and economic and societal impacts. It also says it wants AI developers to compete on safety, not only on capability.

02

Timeline

  1. 2021

    Founded by former OpenAI staff. The first reported raise was about US$ 124 million, in May.

  2. 2022

    Publication of Claude’s first “constitution”, the basis of the technique called Constitutional AI.

  3. Mar. 2023

    First version of Claude, initially only for users approved by the company.

  4. Jul. 2023

    Claude 2, the first model open to the general public.

  5. Sep.–Oct. 2023

    Amazon and Google announce partnerships and investments in Anthropic.

  6. Nov. 2023

    Claude 2.1 expands the context window to 200,000 tokens, about 500 pages.

  7. Mar. 2024

    Claude 3 arrives in three sizes: Haiku, Sonnet and Opus.

  8. Jun. 2024

    Claude 3.5 Sonnet and the launch of Artifacts, for generating and interacting with documents and code.

  9. Oct. 2024

    Computer use, which lets an application use Claude to operate a keyboard and mouse, via the API.

  10. Feb.–May 2025

    Claude 3.7 Sonnet, Claude Code (a command-line coding agent, generally available in May) and Claude 4 (Sonnet 4 and Opus 4), with MCP connectors in the API.

  11. Aug.–Dec. 2025

    Claude in Chrome extension; Agent Skills launched in October and published as an open standard in December.

  12. 2026

    Claude Cowork (January), Claude Design (April), Claude Fable 5 and Claude Science (June), Opus 5 (July), Fable 5.1, Opus 5.5 and Sonnet 5.5 (September) and Claude Haiku 5.5 (October 7).

  13. Oct. 22, 2026

    Anthropic announces it is donating the Agent Skills standard to the Agentic AI Foundation, part of the Linux Foundation, which already hosts the Model Context Protocol (MCP).

03

Products

Claude (apps)

The chat assistant, on web, desktop and mobile, with Pro, Max, Team and Enterprise plans.

API and Claude Platform

Access for developers. The models are also on Amazon Web Services, Google Cloud and Microsoft Azure.

Claude Code

A coding agent that runs in the terminal, launched in 2025. Cowork is the graphical version for people who do not code.

Artifacts

A space to create and review documents, code and prototypes next to the conversation.

Computer use and Claude in Chrome

Features for Claude to operate interfaces and browse the web at the user’s request.

Agent Skills and MCP

Open standards created by Anthropic: skills are folders with instructions and files that teach an agent to perform tasks; MCP connects models to tools and data. Both were donated to the Agentic AI Foundation.

Claude Design and Claude Science

2026 products for visual creation and for scientific research, respectively.

Model families

Haiku (smaller and faster, for volume), Sonnet (balanced) and Opus (most capable). Higher tiers also exist, with restricted access or extra safeguards.

04

Architecture and training

Claude is a transformer-type language model, pretrained to predict the next token in large amounts of text. After pretraining, it goes through tuning with reinforcement learning from human feedback (RLHF) and Constitutional AI.

In Constitutional AI, the model is trained to follow a “constitution”, a set of written principles, so that it is helpful, harmless and honest without relying only on large volumes of human feedback. The first constitution was published in 2022 and the 2023 update listed 75 guidelines; according to reports, the 2026 version has about 23,000 words.

Since Claude 3, models follow the split by size: Haiku, Sonnet and Opus. Haiku 5.5 is the first Haiku with an adjustable “effort” setting (from low to max) and uses an updated tokenizer, which uses slightly more tokens per task, according to the company.

Anthropic also publishes interpretability research. In 2024 it used a “dictionary learning” technique to identify millions of internal features in Claude; in 2025 it reported signs that the model can plan ahead, for example when rhyming verses, and that it processes concepts before converting them into a specific language.

05

What to look at

The declared emphasis is on safety, alignment research and tools for agents (Claude Code, Skills, MCP). Benchmark results quoted in announcements are the company’s own and depend on test conditions.

Like any language model, Claude can make mistakes and invent information. For important decisions, check the source and measure the result on your own tasks. Plans, limits and features vary by account and region.

06

Sources