What Is Anthropic? The Company Behind Claude

Anthropic is the company that makes Claude. It describes itself as “an AI safety and research company” that builds “reliable, interpretable, and steerable AI systems.” It was founded at the beginning of 2021 by seven co-founders, among them Dario Amodei, its chief executive, and Daniela Amodei, its president. It is a Delaware public benefit corporation, with an independent trust that elects part of its board. It sells the Claude apps, Claude Code and the Claude API, and it publishes the policies it says govern how its models are built and released. Everything below comes from Anthropic’s own pages as of September 30, 2026, and each claim says where.

If you came here for the assistant rather than the company, start with what is Claude AI.

Anthropic in its own words

Anthropic’s company page↗ (opens in a new tab) says it “is dedicated to building systems that people can rely on and generating research about the opportunities and risks of AI.” The same page lists seven values the company holds itself to:

• Act for the global good.
• Hold light and shade.
• Be good to our users.
• Ignite a race to the top on safety.
• Do the simple thing that works.
• Be helpful, honest, and harmless.
• Put the mission first.

The pairing in that first description matters for understanding everything else it does. Anthropic presents itself as a safety research lab that also ships a commercial product, and argues the two belong together: its view, set out below, is that safety research has to be done on the most capable systems.

When Anthropic was founded, and by whom

Anthropic’s 2022 announcement of its Series B↗ (opens in a new tab) dates the company “since its founding at the beginning of 2021”, and at that time described a team of around 40 people in San Francisco. Funding rounds and valuations are left out of this page on purpose; they change often, and Anthropic’s newsroom is the place for announced figures.

Anthropic’s leadership page lists seven co-founders:

• Dario Amodei, chief executive, who leads research direction and strategy.
• Daniela Amodei, president and chair of the board.
• Jared Kaplan, chief science officer.
• Tom Brown, chief compute officer.
• Sam McCandlish, chief architect, focused on large-scale model training.
• Chris Olah, who leads interpretability research.
• Jack Clark, head of public benefit, who leads The Anthropic Institute, a think tank announced on March 11, 2026 to study how powerful AI will affect society.

How the company is structured

Anthropic is not a nonprofit. The company page says it “is a Public Benefit Corporation, whose purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.” As Anthropic’s own post on its governance puts it, Delaware law “expressly permits the directors of a PBC to balance the financial interests of the stockholders with the public benefit purpose.”

The unusual part is the Long-Term Benefit Trust↗ (opens in a new tab). Anthropic’s September 19, 2023 post describes it as “an independent body of five financially disinterested members” holding a special class of stock that lets it elect and remove some of the board, phasing in until, “in any event, the Trust will elect a majority of the board within 4 years.” The company page today names three trustees: Neil Buddy Shah, Richard Fontaine and Ben Bernanke. The 2023 design and the current roster do not match on numbers, and Anthropic’s pages do not explain the difference, so read the trust post for the design and the company page for who serves now.

What Anthropic makes

• The Claude apps: the assistant on the web, desktop and mobile, with projects, artifacts, connectors and Cowork. See what is Claude AI.
• Claude Code: an agentic coding tool that works inside a code repository. See what is Claude Code.
• The Claude API and Claude Console: the models for developers, also sold through Amazon Bedrock, Google Cloud and Microsoft Foundry. See the Claude API guide.
• The Claude models themselves: currently Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5, compared in Claude Sonnet vs Opus.
• Free courses through Anthropic Academy, covered in Anthropic Academy.

Individuals use the apps on the Free plan or pay for Pro or Max; businesses buy seats on Team and Enterprise plans or pay for API usage.

The safety approach, as Anthropic describes it

Anthropic’s 2023 essay “Core Views on AI Safety” sets out the reasoning: it argues that safety research has to be done on frontier systems, because many serious problems may only appear in the most capable models. It names research directions including mechanistic interpretability, reverse-engineering what happens inside a network, and scalable oversight. Three published documents turn that view into policy.

The Responsible Scaling Policy

The Responsible Scaling Policy↗ (opens in a new tab), or RSP, is Anthropic’s framework for deciding what safeguards a model needs before it is trained further or released. The current version is 3.4, effective July 8, 2026. It defines capability thresholds, in areas such as chemical, biological, radiological and nuclear (CBRN) weapons and automated AI research, and says that reaching certain thresholds requires upgrading to the ASL-3 Security Standard or the ASL-3 Deployment Standard, ASL standing for AI Safety Level. Anthropic publishes periodic Risk Reports under it, the latest dated August 2026, along with a Frontier Safety Roadmap and a policy for reporting noncompliance without retaliation.

One visible result is that Anthropic keeps some models back. Its Project Glasswing page says the cyber capabilities of its Mythos models carry significant risk if widely distributed, and Claude Mythos 5.1 is offered only by invitation to Glasswing participants, while Fable 5.1, which shares its specifications, is open to all customers.

Claude’s constitution

Claude’s constitution↗ (opens in a new tab) is Anthropic’s written description of the values and behavior it wants Claude to have, and it says the document “plays a crucial role in our training process.” It lists four properties in priority order, to be followed in that order when they conflict: broadly safe, meaning not undermining appropriate human oversight of AI during the current phase of development; broadly ethical; compliant with Anthropic’s guidelines; and genuinely helpful. It is written with Claude as its main audience, which is why it reads more like a letter than a policy.

The Transparency Hub

Anthropic’s Transparency Hub↗ (opens in a new tab) collects model reports for each Claude model, with capabilities, safety evaluations, RSP assessments and links to the full system cards, plus sections on system trust and reporting and on the voluntary commitments it has made. If you want to know what testing a specific model went through before release, this is where to look.

How to check Anthropic’s claims yourself

Most articles about Anthropic repeat each other. The primary documents are short enough to read, and each answers a different question:

• What does the company say it is for? The company page and its seven values.
• Who controls the board over time? The Long-Term Benefit Trust post, then the trustee list on the company page.
• What will stop a dangerous model shipping? The Responsible Scaling Policy and its latest Risk Report.
• How is Claude supposed to behave, and in what order of priority? The constitution.
• What was a particular model tested for? That model’s entry in the Transparency Hub and its system card.
• Which models are current, and when will older ones retire? The models overview and the model deprecations page in Anthropic’s developer docs.
• What does a plan include? Anthropic’s plan and help center pages, which change more often than anything above.

Read them with the same care as any company’s description of itself. They tell you what Anthropic commits to and how it reasons; outside reporting and your own use tell you how that plays out.

The same idea, at the size of a team

Anthropic’s approach comes down to writing the rules down before the model acts, and saying who decides. A team using Claude, or any other assistant, needs a small version of the same thing. On fenbs, a task board that AI assistants join over MCP, the Decisions and rules page holds it: a rule is a decision that holds from now on, the decider is always a person, and every connected assistant reads the rules first through fenbs_get_context before it touches a task. History then records which person’s assistant changed what. fenbs does not enforce anything inside Claude; it gives Claude your rules to read and keeps a record of what it did with them.