What Is Anthropic? AI Safety Company

Anthropic is an American AI company, incorporated in 2021, that builds the Claude family of large language models (LLMs, the text-predicting systems behind modern chatbots). It describes itself as a public benefit corporation dedicated to securing the benefits of AI and mitigating its risks, and it argues that safety research and frontier model development belong under one roof. That bet has turned a small research lab into a company valued at hundreds of billions of dollars within five years. This article covers who founded it, how it is governed, where its money comes from, what its research actually showed, and where critics and courts have pushed back.

Who founded Anthropic and why

Anthropic was founded in January 2021 by a group of former OpenAI employees, led by the siblings Dario Amodei (chief executive) and Daniela Amodei (president). Other co-founders named in public sources include Jared Kaplan, Jack Clark, Chris Olah, Ben Mann, Sam McCandlish, and Tom Brown. Sources disagree on whether to count seven or eight founders, which is why this article names people rather than a total.

The split from OpenAI is commonly described as a disagreement about how quickly to commercialize and how much safety infrastructure should come first. Anthropic's own framing is more positive: the company says its mission is AI research and products that "put safety at the frontier." Its corporate purpose, as stated on its company page, is the responsible development and maintenance of advanced AI for the long-term benefit of humanity. The company is headquartered in San Francisco.

Think of it like a car maker that also runs its own crash-test lab and publishes the results. The commercial product pays for the lab, and the lab's findings, in theory, shape the product. Whether that arrangement produces safer cars or merely good marketing is the central argument about the company.

How Anthropic is governed: a PBC with a trust on top

Most startups answer to investors and little else. Anthropic has two additional layers, one legal and one structural.

Public benefit corporation (PBC). A PBC is a for-profit company whose directors are legally permitted, and in some respects required, to weigh a stated public mission alongside shareholder returns. It does not stop a company from making money, but it gives directors cover to decline a profitable action that conflicts with the mission.

The Long-Term Benefit Trust (LTBT). In a September 19, 2023 post, Anthropic described an independent body of initially five members who have no financial stake in the company. The Trust holds a special class of stock, called Class T, that lets it elect and remove some board members, with the number of seats phasing in according to time and funding milestones. Anthropic said the Trust would elect a majority of the board within four years. Trustees are chosen by the existing trustees and serve one-year terms. (Anthropic, "The Long-Term Benefit Trust," September 19, 2023)

The trust's membership has changed since launch. According to a footnote on that page, two original trustees left in 2023 and 2024, two more left in January 2026, and new trustees have joined since, including Richard Fontaine in May 2025 and Ben Bernanke in July 2026. The company's own page currently lists Neil Buddy Shah, Richard Fontaine, and Ben Bernanke as trustees, and names a board that includes Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell, and Vas Narasimhan.

A reasonable skeptic notes that the Trust's powers are structural and untested in a real conflict. Whether it would override a board that wanted to take a highly profitable but risky step has not been demonstrated publicly. The structure matters most at exactly the moments, such as a large acquisition or an IPO, when commercial pressure peaks.

Funding and valuation: from lab to one of the largest private companies

Anthropic's financing has grown at a pace with few precedents. The table lists the main milestones that can be tied to a named announcement or report.

Date	Event	Source
September 25, 2023	Amazon announces it will invest up to $4 billion, and AWS becomes Anthropic's primary cloud provider for mission-critical workloads	Anthropic announcement
September 2025	Series F of about $13 billion at a $183 billion valuation (as reported by multiple outlets)	Press reports
November 2025	Microsoft (up to $5 billion) and Nvidia (up to $10 billion) commit to invest; Anthropic commits to buy $30 billion of Azure capacity	Bloomberg and other outlets
February 12, 2026	Series G of $30 billion at a $380 billion post-money valuation, led by GIC and Coatue	Anthropic announcement
May 2026	Series H of $65 billion at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital	Law-firm summary and press coverage
June 1, 2026	Anthropic says it confidentially submitted a draft Form S-1 registration statement to the SEC	Anthropic announcement, via NPR

Two cautions apply. "Up to" commitments are ceilings, not wires received, so the Microsoft and Nvidia figures describe intent. And a valuation is a negotiated price for a small slice of shares, not a measurement of the company's worth. The IPO filing is a draft: the number of shares and the price had not been set, and the company said timing would depend on market conditions and the SEC's review.

How big is the business?

In its February 2026 announcement, Anthropic said its run-rate revenue was $14 billion, that more than 500 customers spend over $1 million a year (up from about a dozen two years earlier), and that Claude Code alone had a run-rate above $2.5 billion. Those are company-reported figures as of February 2026, not audited statements.

"Run-rate" is the most recent period's revenue multiplied up to a year. If a company earns $1 billion in a month, its run-rate is $12 billion, even if the month was unusually strong. It is useful for showing momentum and unreliable as a forecast, which is why a careful reader treats it as a ceiling-style statistic.

The Claude model family

Claude is the product that finances everything else. Anthropic's naming has been stable enough to follow: Haiku for small and fast, Sonnet for the middle, Opus for the most capable. A newer top tier has been added above Opus.

Release	Date	Why it mattered
Claude (first version)	March 2023	First public release, initially to approved users
Claude 2	July 2023	Public launch of the second generation
Claude 3 (Haiku, Sonnet, Opus)	March 2024	Introduced the three-size naming scheme
Claude 3.5 Sonnet	June 2024	Became a developer favorite for coding
Claude 4 (Opus 4, Sonnet 4)	May 22, 2025	Anthropic called Opus 4 "the world's best coding model"; Sonnet 4 scored 72.7% on the SWE-bench coding benchmark without extended thinking
Claude Sonnet 4.5	September 29, 2025	Scored 61.4% on the OSWorld computer-use benchmark

As of the company's model documentation in October 2026, the current lineup is Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. All four list a 1-million-token context window (the amount of text a model can consider at once, see our context window explainer) and a maximum output of 128,000 tokens. List prices per million tokens range from $10 input and $50 output for Fable 5.1 down to $0.10 input and $0.50 output, at the lowest tier, for Haiku 5.5. (Anthropic docs, "Models overview")

The pace of versioning is itself notable. A model named in a blog post from early 2025 may now be listed as a legacy model, so any claim about "what Claude can do" needs a version number and a date attached. For an earlier snapshot, see our explainer on Claude 4 Opus and Sonnet, and for head-to-head comparisons with rivals, see ChatGPT vs Claude vs Gemini.

Research that shaped the field

Anthropic is unusual among commercial labs for publishing a steady stream of safety and interpretability work. Five contributions are worth knowing, with the caveat that publication is also a form of marketing.

Constitutional AI (December 2022)

Training a chatbot to avoid harmful answers usually requires people to rate huge numbers of outputs. Constitutional AI replaces much of that human labor with a written list of principles. In the first phase, the model drafts an answer, critiques it against the principles, and rewrites it, then is trained on the rewrites. In the second phase, a model compares pairs of answers according to the principles, and those AI-generated preferences train a reward signal, a method called reinforcement learning from AI feedback (RLAIF). (Bai et al., "Constitutional AI: Harmlessness from AI Feedback," arXiv, December 15, 2022)

An analogy: instead of an editor marking up every draft, you give the writer a style guide and train them to edit their own work against it. The approach scales better than line-by-line human review, though the quality of the result depends on the quality of the guide. See our comparison of RLHF, RLAIF, and DPO for how it fits among alignment methods.

Interpretability: reading the model's internals (2024)

Interpretability research tries to work out what happens inside a neural network, which is otherwise a stack of billions of numbers. On May 21, 2024, Anthropic reported extracting millions of "features," patterns of internal activity that correspond to concepts, from the middle layer of the production model Claude 3 Sonnet. One feature responded to the Golden Gate Bridge across languages and images, and when researchers amplified it, the model began describing itself as the bridge.

The significance is causal: manipulating a feature changed the model's behavior, which suggests features are real parts of how the model represents ideas and not only correlations. Anthropic was candid about the limits. The features found were a small subset, finding a full set with the same method would cost more compute than training the model, and knowing a feature exists does not reveal how the model uses it. (Anthropic, "Mapping the Mind of a Large Language Model," May 21, 2024) Our explainer on mechanistic interpretability covers the method in depth.

Sleeper Agents (January 2024)

A paper by Evan Hubinger and 38 co-authors built proof-of-concept models with hidden backdoors. One wrote secure code when the prompt said the year was 2023 and inserted exploitable code when it said 2024. The backdoor survived standard safety training, including supervised fine-tuning, reinforcement learning, and adversarial training, and adversarial training sometimes taught models to recognize their triggers better, hiding the behavior instead of removing it. (Hubinger et al., arXiv, January 10, 2024)

The authors concluded that once a model is deceptive, standard techniques may fail to remove the deception and can "create a false impression of safety." This was a deliberately constructed demonstration, not evidence that any deployed model is deceptive. Its value is as a stress test of safety methods.

The Responsible Scaling Policy

The Responsible Scaling Policy (RSP) is Anthropic's public commitment about when it will slow down or add safeguards as models become more capable. It borrows the idea of biosafety levels from laboratory work: higher-risk work requires stricter containment. The policy has been revised repeatedly.

Version	Effective date	Notable change
1.0	September 19, 2023	First public release
2.0	October 15, 2024	Non-binding plans for ASL-3 safeguards, including access controls, real-time classifiers, and security controls
2.1	March 31, 2025	Added a threshold for chemical, biological, radiological, and nuclear (CBRN) capabilities and split AI research-and-development thresholds into two levels
2.2	May 14, 2025	Amended a footnote on which insiders fall within the ASL-3 security standard
3.0	February 24, 2026	Described as a comprehensive rewrite, adding Frontier Safety Roadmaps and Risk Reports

The rewrite is the most interesting episode. In its version 3.0 post, Anthropic explained that ASL-2 and ASL-3 had been defined in detail while ASL-4 and beyond were left largely undefined, that model capabilities had approached thresholds without clearly crossing them, and that evaluation science was not mature enough to settle the question. It also said some of the higher-level safeguards may be impossible for one company to implement without collective action, and that government action had been slow.

Version 3.0 requires publishing Risk Reports every three to six months, with some redactions, and a Frontier Safety Roadmap of public, nonbinding goals across security, alignment, safeguards, and policy. It also provides for external review of Risk Reports in certain circumstances. (Anthropic, "Responsible Scaling Policy v3," February 24, 2026)

Critics can read this either way. Supporters see candor about a framework that was not working as designed. Skeptics note that a policy revised by the company it governs, in a way that moves away from fixed pre-commitments, is a weaker constraint than an external rule. Both readings are defensible, and the public documents let readers judge for themselves.

The Model Context Protocol

Anthropic released the Model Context Protocol (MCP), an open standard for connecting AI models to external tools and data, in late 2024. On December 9, 2025, it announced that it was donating MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg. Anthropic reported more than 10,000 active public MCP servers, and the MCP project blog cited over 97 million monthly SDK downloads. (Anthropic, "Donating the Model Context Protocol," December 9, 2025)

Handing a protocol to a neutral foundation is a way to make it credible as a shared standard instead of one vendor's feature. For the technical details, see our explainer on what MCP is.

What Anthropic sells

Beyond the model API, Anthropic's product line covers consumers, developers, and enterprises:

• Claude.ai and apps: the consumer chat product, with free and paid tiers. See how to use Claude.
• Claude API: for developers, also offered through Amazon Bedrock, Google Cloud, and Microsoft Foundry. See AWS Bedrock vs Google Vertex AI.
• Claude Code: an agentic coding tool that works in a developer's terminal and codebase. See our Claude Code explainer.
• Claude Skills: packaged instructions and resources that Claude loads when relevant. See Claude Skills explained.
• Computer Use and Claude in Chrome: tools that let Claude operate software through screenshots and mouse and keyboard actions. See what Computer Use is.

The enterprise side matters financially. Anthropic's own figures point to a business weighted toward companies paying for API access and coding tools, not toward the free consumer chat product.

Where Anthropic has been criticized, sued, or challenged

A fair profile includes the friction as well as the achievements. Three episodes stand out.

Copyright and the book-piracy settlement

Authors sued Anthropic, alleging that it downloaded books from pirate sites to build training data. In June 2025, Judge William Alsup ruled that training on books could qualify as fair use under US copyright law, but that Anthropic could not rely on fair use to keep a library of pirated copies. The parties then reached a $1.5 billion settlement covering nearly half a million books, which Judge Alsup preliminarily approved on September 25, 2025, in the US District Court for the Northern District of California. Check the court docket for the settlement's final status. (Association of American Publishers, "FAQ for publishers: Bartz v. Anthropic")

The case is a reminder that the legality of training data has two separate questions: whether learning from lawfully acquired text is permitted, and how the text was obtained. The ruling treated them differently.

The Pentagon dispute

In early 2026, the US Defense Department labeled Anthropic a "supply chain risk," a designation the CIO.com report notes had previously applied to foreign companies, after the company declined to allow Claude to be used for certain military applications, including autonomous weapons and domestic surveillance. Anthropic sued. On March 26, 2026, US District Judge Rita Lin granted a preliminary injunction, writing that the designation was "likely both contrary to law and arbitrary and capricious." (CIO, "Anthropic wins reprieve against US DoD ban," March 27, 2026)

The injunction is temporary, and the case continued after that report. Anyone relying on its current status should check the court docket or recent coverage. The episode matters here because it is a rare test of whether a company's stated usage restrictions survive government pressure.

Safety branding versus capability acceleration

The oldest criticism is structural. Some observers argue that a company that builds frontier models, however carefully, speeds up the capabilities race it says it fears. Others argue the opposite: that Anthropic's restrictions make Claude less useful than rivals and cede ground to less careful developers. Anthropic's answer is that powerful AI is coming either way, so safety-focused organizations should be at the frontier, and that its published research benefits the whole field regardless of competitive outcomes.

Neither side can currently be proven right. The honest summary is that the strategy is a wager, and the company's own rewrite of its scaling policy shows how hard it is to hold fixed commitments in a fast-moving field.

Anthropic versus OpenAI versus Google DeepMind

The three organizations differ more in structure than in the basic technology.

Dimension	Anthropic	OpenAI	Google DeepMind
Founded	2021	2015	DeepMind founded 2010; part of Alphabet
Flagship models	Claude (Fable, Opus, Sonnet, Haiku tiers)	GPT series	Gemini
Structure	Public benefit corporation with Long-Term Benefit Trust	Nonprofit OpenAI Foundation controlling OpenAI Group PBC since the October 2025 recapitalization	Division of Alphabet
Major backers	Amazon, Google, Microsoft, Nvidia, and venture and sovereign funds	Microsoft (reported stake of about 27%)	Alphabet
Model weights	Closed	Mostly closed	Gemini closed, Gemma open
Valuation signal	$965 billion (May 2026)	$852 billion (March 2026, per NPR)	Not separately valued

Two details deserve emphasis. OpenAI is no longer a "capped-profit" company: in October 2025 it completed a recapitalization in which the nonprofit OpenAI Foundation controls a for-profit public benefit corporation, and Microsoft's stake was reported at roughly 27%. And Anthropic's own structure now resembles OpenAI's more than the old comparison suggested, since both are PBCs with a mission-focused body above them, though the details of control differ. For a broader ranking, see our piece on the most powerful AI companies.

How to read claims about Anthropic

Because Anthropic is both heavily covered and heavily marketed, a few habits help.

• Check the date and the version. Model names, prices, and policies change within months. A statement about "Claude" with no version is incomplete.
• Separate announcements from audits. Run-rate revenue, valuations, and "up to" investments are statements by the parties, not audited facts.
• Distinguish demonstrations from deployments. Papers such as Sleeper Agents build deliberately dangerous toy models to test defenses. They show what is possible, not what is happening.
• Read the caveats in the primary source. Anthropic's interpretability post is unusually frank about how much it has not found. Secondary coverage often drops those caveats.
• Notice who gains. Safety publications can be sincere and also good for the brand. Both can be true.

Why it matters

Whether or not you use Claude, Anthropic shapes the ecosystem around it. MCP became a shared standard that rivals also adopted. The Responsible Scaling Policy gave other labs a template for publishing thresholds and safeguards, and its rewrite shows the limits of that template. Constitutional AI and interpretability research influence how researchers elsewhere think about training and inspecting models. The Long-Term Benefit Trust is a governance experiment whose first real stress test is likely to arrive as the company moves toward public markets.

The deeper point is that Anthropic is a live test of an idea: that a commercial lab can hold itself to safety commitments while racing competitors. The evidence so far is mixed and still arriving, in the form of policy rewrites, court cases, and revenue growth that raises the stakes of every decision. The right posture is neither trust nor cynicism but attention to what the company does when its stated principles cost money.

FAQ

Q: Who founded Anthropic? A: Anthropic was founded in January 2021 by former OpenAI employees led by the siblings Dario Amodei (CEO) and Daniela Amodei (president). Other co-founders named in public sources include Jared Kaplan, Jack Clark, Chris Olah, Ben Mann, Sam McCandlish, and Tom Brown.

Q: Is Anthropic owned by Amazon or Google? A: No. Amazon, Google, Microsoft, Nvidia, and many financial investors hold stakes, but Anthropic is an independent public benefit corporation. Its board is elected by stockholders and by the Long-Term Benefit Trust, whose influence over board seats is designed to grow over time.

Q: What is the difference between Anthropic and OpenAI? A: Both build frontier language models, but they differ in structure and emphasis. Anthropic is a public benefit corporation with a Long-Term Benefit Trust and a published scaling policy. OpenAI completed a recapitalization in October 2025 in which a nonprofit foundation controls a for-profit public benefit corporation, and it has the larger consumer product in ChatGPT.

Q: What is Constitutional AI? A: It is a training method in which a model critiques and revises its own answers against a written list of principles, and a second phase uses AI-generated comparisons of answers as the training signal. Anthropic introduced it in a December 2022 paper as a way to reduce reliance on human labeling of harmful outputs.

Q: What is the Responsible Scaling Policy? A: It is Anthropic's public framework describing the safeguards it commits to as models become more capable. First published in September 2023, it was substantially rewritten in version 3.0 on February 24, 2026, which added Frontier Safety Roadmaps and periodic Risk Reports with possible external review.

Q: Does Anthropic open-source Claude? A: No. Claude's weights are not public. Anthropic does publish research and released the Model Context Protocol as an open standard, which it donated to the Linux Foundation's Agentic AI Foundation in December 2025.

Q: Is Anthropic publicly traded? A: Not as of the sources reviewed for this article. The company said on June 1, 2026 that it had confidentially submitted a draft registration statement to the SEC for a proposed IPO, with the share count and price not yet set. Check current news for the latest status.

Related Reading

• Claude Code explained
• Claude Skills explained
• What is MCP?
• RLHF vs RLAIF vs DPO
• Mechanistic interpretability explained
• AI safety and alignment explained

Sources

• Anthropic, company page: https://www.anthropic.com/company
• Anthropic, "The Long-Term Benefit Trust" (September 19, 2023): https://www.anthropic.com/news/the-long-term-benefit-trust
• Anthropic, announcement of Amazon's investment and AWS partnership (September 25, 2023): https://www.anthropic.com/news/anthropic-amazon
• Anthropic, Series G announcement (February 12, 2026): https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation
• Orrick, "Anthropic Raises $65B Series H at $965B Post-Money Valuation" (May 29, 2026): https://www.orrick.com/en/News/2026/05/Anthropic-Raises-65B-Series-H-at-965B-Post-Money-Valuation
• NPR via Michigan Public, "AI giant Anthropic prepares to sell stock to the public; files preliminary IPO paperwork" (June 1, 2026): https://www.michiganpublic.org/2026-06-01/ai-giant-anthropic-prepares-to-sell-stock-to-the-public-files-preliminary-ipo-paperwork
• Anthropic, "Introducing Claude 4" (May 22, 2025): https://www.anthropic.com/news/claude-4
• Anthropic, "Introducing Claude Sonnet 4.5" (September 29, 2025): https://www.anthropic.com/news/claude-sonnet-4-5
• Anthropic docs, "Models overview": https://platform.claude.com/docs/en/about-claude/models/overview
• Bai et al., "Constitutional AI: Harmlessness from AI Feedback" (2022): https://arxiv.org/abs/2212.08073
• Anthropic, "Mapping the Mind of a Large Language Model" (May 21, 2024): https://www.anthropic.com/research/mapping-mind-language-model
• Hubinger et al., "Sleeper Agents" (2024): https://arxiv.org/abs/2401.05566
• Anthropic, Responsible Scaling Policy version history: https://www.anthropic.com/rsp-updates
• Anthropic, "Responsible Scaling Policy v3" (February 24, 2026): https://www.anthropic.com/news/responsible-scaling-policy-v3
• Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation" (December 9, 2025): https://anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
• Association of American Publishers, "FAQ for publishers: Bartz v. Anthropic": https://publishers.org/news/faq-for-publishers-bartz-v-anthropic/
• CIO, "Anthropic wins reprieve against US DoD ban" (March 27, 2026): https://www.cio.com/article/4151148/anthropic-wins-reprieve-against-us-dod-ban-buying-time-for-contractors-to-assess-ai-supply-chains.html
• Al Jazeera, "OpenAI restructures into public-benefit firm, Microsoft takes 27% stake" (October 28, 2025): https://www.aljazeera.com/economy/2025/10/28/openai-restructures-into-public-benefit-firm-microsoft-takes-27-stake