Anthropic

Anthropic is an American artificial intelligence company based in San Francisco, founded in 2021 by a group of researchers who had worked at OpenAI. It builds the Claude family of language models and sells access to them through a developer platform and a consumer subscription, and it has built its public identity around the argument that safety research and commercial success are compatible 1.

The company is a public benefit corporation, a structure that obliges directors to weigh its mission alongside shareholder returns. Its governance includes a trust that appoints a minority of the board, an arrangement intended to make the mission difficult to overturn 2.

Anthropic is smaller than OpenAI by revenue and larger than most of its other competitors. It is also the most explicit of the frontier laboratories about the techniques it uses to shape model behaviour, and the most exposed to the argument that those techniques are as much marketing as engineering 3.

Most of the founders left OpenAI between 2020 and 2021 after disagreements about the direction of that laboratory, and several have said publicly that they wanted to build a company in which safety work would not be subordinated to product deadlines. Dario Amodei, formerly OpenAI's vice-president of research, became chief executive; his sister Daniela Amodei, previously at Stripe, became president; Tom Brown, Chris Olah, Sam McCandlish, Jack Clark and Jared Kaplan completed the founding group 3.

Anthropic incorporated as a public benefit corporation, a Delaware form that requires directors to balance mission and profit. In 2023 it added a Long-Term Benefit Trust, whose members are chosen for expertise rather than shareholding and which elects a minority of the board. The company argues that this protects its charter; critics note that a minority of seats is not control 2.

Early work concentrated on interpretability and on alignment methods that could be tested at scale. The company had no consumer product for its first two years, and its revenue came from research contracts and early API access. That changed with Claude, released in March 2023, which made Anthropic a commercial competitor rather than a research group with a point of view 1.

Claude was first released in March 2023, positioned around long documents and careful refusals. Claude 2, in July 2023, extended the context window to 100,000 tokens — roughly a long novel — and made document analysis a core use case. The Claude 3 family, in March 2024, introduced the naming scheme still in use: Haiku for speed and low cost, Sonnet for general work, Opus for the hardest tasks 3.

Claude 3.5 Sonnet, released in June 2024, was the model that made Anthropic a serious competitor in software engineering. It was followed by computer use, a capability that let the model operate a desktop through screenshots and clicks, and by Claude 3.7 Sonnet in February 2025, which introduced a visible extended thinking mode that users could budget for 3.

Claude Opus 4 and Sonnet 4 arrived in May 2025, with improvements aimed at long-running coding tasks and agents. Opus 4.1 followed in August, and Sonnet 4.5 in September. The company has kept the family small and the version numbers legible, a contrast with competitors who release several models a year under overlapping names 4.

Distribution runs through three channels: Anthropic's own API and applications, Amazon Bedrock, and Google Cloud's Vertex AI. The multi-cloud arrangement reflects the fact that Anthropic's two largest investors are also its two largest hosting partners 3.

Anthropic's best-known technical contribution is Constitutional AI, described in a 2022 paper. The method gives the model a written set of principles and trains it to critique and revise its own responses against them, using AI feedback rather than large amounts of human labelling of harmful text. The result is a model that can explain a refusal by reference to a stated rule rather than an opaque preference 5.

The second pillar is the Responsible Scaling Policy, adopted in 2023. It defines AI Safety Levels, each tied to capability thresholds in areas such as biological weapons, cyber offence and autonomous replication, and commits the company to stronger security and evaluation before passing each threshold. The policy is a self-imposed regime: the company sets the thresholds, runs the evaluations and decides when they are met 2.

A third strand is interpretability. Anthropic's researchers have mapped circuits inside small models, extracted interpretable features from larger ones, and argued that this work is a prerequisite for supervising systems whose reasoning cannot be read off their outputs. The company has also funded work on model welfare, an unusual line of research that treats the question of model experience as one worth taking seriously 1.

The combination is coherent, and it is also a commercial position. Being the laboratory that talks most about safety attracts enterprise customers who need a defensible vendor, and it attracts researchers who would otherwise stay in academia.

Anthropic raised early rounds from individuals and funds including Jaan Tallinn, Eric Schmidt and Dustin Moskovitz, then took strategic money from the two clouds it depends on. Google invested about 300 million dollars for roughly a tenth of the company in October 2023 and later committed up to two billion more; Amazon invested 1.25 billion dollars in September 2023, added 2.75 billion in March 2024, and a further four billion in November 2024, for a total of eight billion 3.

The company's valuation rose steeply through 2025. A Series E in March raised 3.5 billion dollars at a post-money valuation of 61.5 billion. A Series F in September, led by ICONIQ and co-led by Fidelity and Lightspeed, raised 13 billion at 183 billion, with sovereign funds, pensions and asset managers participating alongside venture firms 6.

The revenue disclosed alongside that round was unusually specific for a private company: run-rate revenue of about one billion dollars at the start of 2025, more than five billion by August, over 300,000 business customers, and Claude Code alone above 500 million in run-rate revenue. These are company figures, not audited accounts 6.

Independence is the unresolved question. Amazon and Google together hold stakes that would be material in any public listing, both host Claude, and both sell their own competing models. Anthropic has so far avoided exclusive arrangements, and its investors have accepted that.

Anthropic sells tokens through an API priced by model tier, and seats through consumer and team subscriptions. Its commercial pitch emphasises reliability on long tasks, low rates of fabricated citations, and behaviour that is easy to constrain — claims that matter most to customers in law, finance, software and life sciences, where an unverifiable answer is worse than no answer 1.

Claude Code, released as a research preview in February 2025 and fully launched in May, became the company's fastest-growing product. It runs in a terminal, reads a repository, edits files and runs tests, and it turned Anthropic from a model vendor into a tools vendor. The company has since extended the same approach to other surfaces, including a desktop application and integrations with development environments 6.

Anthropic also published the Model Context Protocol in November 2024, a specification for connecting models to external tools and data. It was adopted by other laboratories during 2025, which gave Anthropic an unusual form of influence: a small company setting a standard that larger competitors implemented 1.

Its enterprise business is concentrated in coding and in document-heavy workflows. That focus is a strength and a limit; Anthropic has less consumer distribution than OpenAI or Google, and no search engine, office suite or mobile operating system of its own.

The most consequential dispute is a copyright class action brought in August 2024 by Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, who alleged that Anthropic had downloaded pirated copies of their books and used them for training. In June 2025 Judge William Alsup granted summary judgment for Anthropic on the use of lawfully acquired books, holding that training was transformative, but allowed the claims over pirated copies to proceed 7.

Anthropic then settled. It agreed to pay at least 1.5 billion dollars plus interest for 482,460 works, roughly 3,000 dollars per book, to destroy the pirated datasets, and to accept releases covering conduct up to August 2025. The settlement was preliminarily approved in September 2025 and received final approval in 2026, with counsel fees cut to 6.8 per cent; about 350 rightsholders opted out 7.

The case turned on Anthropic's own disclosures. The company had described its data pipeline in order to argue fair use, and that description established where the copies came from. The ruling left the larger question — whether training on copyrighted text is lawful — unsettled, because the judge's fair-use finding applied to books the company had bought or scanned rather than to the pirated set 7.

A second dispute concerned consumer data. In August 2025 Anthropic said it would train future models on consumer chats and coding sessions unless users opted out by 28 September, and would extend retention to five years for those who did not. The company presented the change as a way to improve models on real usage; critics described it as a default-on consent regime introduced with little notice 8.

Whether Anthropic is meaningfully safer than its competitors is not a question that current evidence can settle. The company publishes more about its methods than most, submits models to external evaluators, and has declined to release systems it judged too capable; it also funds the evaluations it cites, sets its own thresholds, and sells the models whose risks it assesses 2.

A second disagreement concerns the copyright ruling. Anthropic and its supporters read Judge Alsup's decision as a broad endorsement of training as fair use. Plaintiffs and commentators read the same decision as narrower, because it turned on how the books were acquired and left open the question of market harm from generated text. Both readings draw on the same paragraphs 7.

The governance structure is contested in a quieter way. The Long-Term Benefit Trust was designed to hold the mission against shareholder pressure, but it elects only a minority of directors, and its members are appointed by the company. Whether that constitutes meaningful oversight or a reputational device depends on assumptions that the company has not tested publicly 2.

Finally, the revenue figures that underpin the valuation are run-rate estimates disclosed in a fundraising announcement. They are the best numbers available, and they are not audited.

Anthropic is the second-largest frontier laboratory by revenue and the strongest in enterprise software work. Its models are the default alternative to OpenAI's in coding tools, its safety research is cited widely, and its protocol work has given it influence beyond its size 1.

Its weaknesses are structural. It has no consumer distribution channel comparable to a search engine or an office suite, it depends on two investors who are also competitors, and its safety positioning invites scrutiny that its rivals avoid. The 2026 copyright settlement removed a large liability but established a price for data that other laboratories will have to reckon with 7.

For the encyclopedia, Anthropic matters as the clearest test of a proposition: that a company can make safety a product attribute rather than a constraint, and be valued at 183 billion dollars while doing so. Whether the proposition holds is a question for the record, not for the company's own account of itself 6.