Anthropic

• 1. Introduction

  ›

  Anthropic finished training its first AI model in the summer of 2022 and then sat on it for months. The company wanted to avoid triggering what it described as a potentially hazardous race toward increasingly powerful AI systems. So Claude, as the model was called, waited for internal safety testing to run its course. It spent months in review before anyone outside the company could use it.

  Seven former OpenAI employees founded Anthropic in January 2021, among them siblings Dario and Daniela Amodei. Dario had served as OpenAI's Vice President of Research. They departed over what the company describes as directional differences. Their departure planted a question that has followed the company ever since. Can a commercial AI lab make safety its genuine first priority? One answer Anthropic pursued was to write down exactly what it wanted Claude to believe.

• 2. A Constitution for Claude

  ›

  The technique Anthropic developed to shape Claude's behavior has a formal name: Constitutional AI. Within this framework, a document called a "constitution" describes how the model should act. The model then evaluates its own outputs against those principles and adjusts accordingly.

  Claude 2's constitution drew from sources as varied as the 1948 Universal Declaration of Human Rights and Apple's terms of service. One rule taken directly from the UN Declaration instructs Claude to choose responses that "most support and encourage freedom, equality and a sense of brotherhood." The decision to root an AI's behavior in human rights documents was a deliberate philosophical choice.

  A philosopher named Amanda Askell works specifically on Claude's character. Alongside that work, Anthropic runs a research program devoted to understanding the internal mechanics of its own models. Using a compute-intensive technique called dictionary learning, researchers identified millions of patterns of neural activation inside Claude in 2024. One of those patterns corresponded to the concept of the Golden Gate Bridge.

  In March 2025, Anthropic published research on how multilingual language models process meaning. The findings suggested that these models partially work in a shared conceptual space before converting output into a specific language. The same research found evidence that the models sometimes plan ahead. When writing poetry, Claude appears to identify potential rhyming words before constructing the line that will end with one of them. Whether that planning behavior extended to more consequential tasks was the kind of question the company's interpretability program was built to pursue.

• 3. A Model Called Claude

  ›

  Some employees regard the name Claude as a reference to the mathematician Claude Shannon. Shannon formalized the theory of information in the mid-twentieth century. That possible origin fits a company that presents itself as research-first.

  The first two versions, Claude and Claude Instant, appeared in March 2023 but only for users Anthropic had approved in advance. The fully public release came in July 2023 with Claude 2. By March 2024, Anthropic offered three tiers under the Claude 3 banner: Opus, Sonnet, and Haiku, in descending order of performance. Claude 3.5 Sonnet followed in June 2024. Claude 4 Opus and Sonnet arrived in May 2025, and in April 2026, the company announced a model called Mythos.

  Anthropic makes Claude available through a web interface, an API, Amazon Bedrock, an iOS app, and both Mac and Windows desktop applications. Its Claude Code product is a command-line AI agent designed for software developers. A graphical alternative called Cowork targets non-technical users, enabling them to automate tasks like reading, editing, and writing files without programming knowledge.

  The term "vibe coding" entered technology's vocabulary in February 2025, coined by AI researcher Andrej Karpathy. It describes an approach in which users describe what they want in plain language and let an AI agent handle the code. The arrival of Claude Code helped accelerate the practice across the industry.

  In September 2025, Anthropic published a report on how businesses actually use the model. Three-quarters of companies working with Claude were delegating complete tasks to it, rather than treating it as a collaborative tool. CEO Dario Amodei had predicted earlier in the year that AI would eliminate many white-collar jobs. Entry-level roles in finance, law, and consulting were among those he named. That pace of adoption had drawn investors to Anthropic at a scale the AI industry had rarely seen.

• 4. The Capital Behind Claude

  ›

  In May 2021, just four months after the company's founding, Anthropic raised $124 million. That initial round foreshadowed a fundraising trajectory without clear precedent in the technology industry.

  The first large round came in April 2022, when Anthropic announced $580 million in total funding. That sum included $500 million from FTX, the cryptocurrency exchange run by Sam Bankman-Fried, who would later be convicted of fraud. Amazon entered as a partner in September 2023 with an initial investment of $1.25 billion and a total commitment of $4 billion. The remaining $2.75 billion arrived in March 2024. Amazon then doubled its total stake to $8 billion with an additional $4 billion investment in November 2024.

  Google joined the investor roster in October 2023 with a $500 million stake and a commitment to add $1.5 billion more over time. In March 2025, Google agreed to invest another $1 billion in the company.

  A Series E in March 2025 raised $3.5 billion, valuing Anthropic at $61.5 billion after the investment. Lightspeed Venture Partners led that round. A Series F in September 2025 brought in $13 billion, pushing the post-money valuation to $183 billion. That round was co-led by Iconiq Capital, Fidelity Management & Research, and Lightspeed Venture Partners, with participation from the Qatar Investment Authority.

  On the 31st of December 2025, Anthropic signed a term sheet for a $10 billion round. It was led by Coatue and GIC, at a valuation of $350 billion. On the 12th of February 2026, the company announced a Series G of $30 billion, setting the post-money valuation at $380 billion. Anthropic's estimated value reached $965 billion, and the company remained privately held. In May 2026, The New York Times reported that Anthropic was considering filing for an initial public offering. The company was targeting a market debut in the fall of 2026. That prospective debut would arrive after years of legal scrutiny over how Claude had been trained.

• 5. What Claude Read

  ›

  On the 18th of October 2023, music publishers including Concord, Universal, and ABKCO filed a lawsuit against Anthropic. Their complaint alleged systematic infringement of copyrighted song lyrics used to train Claude. Plaintiffs sought up to $150,000 for each work they claimed had been infringed. Among the examples cited were lyrics from Katy Perry's "Roar" and Gloria Gaynor's "I Will Survive," which Claude had reproduced. On the 16th of January 2024, Anthropic responded that the publishers had not been unreasonably harmed and that the cited instances were merely bugs.

  A class-action lawsuit from authors followed in August 2024. Three of the named participants were Kirk Wallace Johnson, Andrea Bartz, and Charles Graeber. They alleged that Anthropic had fed its models with pirated copies of their books. On the 23rd of June 2025, a federal district court in California ruled on the matter. It found that training on legitimately purchased digital copies of the plaintiffs' works constituted fair use. But it ruled that training on pirated library copies did not. Those claims were ordered to trial. In September 2025, Anthropic agreed to pay $1.5 billion to settle the case, roughly $3,000 per book plus interest. The proposed settlement, pending a judge's approval, was described as the largest copyright resolution in U.S. history.

  Court filings unsealed in January 2026 revealed the existence of what the company internally called Project Panama. A planning document described it as Anthropic's "effort to destructively scan all the books in the world." The company purchased millions of used books from retailers including Better World Books. It cut off the spines, scanned the pages, and recycled the paper. Tom Turvey, who had helped create Google Books, was hired to lead the operation. The planning document noted that Anthropic did not "want it to be known that it was working on this." Judge William Alsup later ruled that destroying legally purchased books to scan them constituted fair use. The pirated copies issue had already been resolved by the settlement.

  Reddit sued Anthropic in June 2025, alleging the company had trained its models on users' personal data without their consent. The lawsuit claimed Anthropic had violated Reddit's user agreement. Even as those cases wound through the courts, Claude was being deployed in classified settings that raised questions well beyond copyright.

• 6. Claude at War

  ›

  In November 2024, Anthropic partnered with Palantir and Amazon Web Services to make Claude available to U.S. intelligence and defense agencies. By June 2025, the company had released a dedicated variant called Claude Gov. Ars Technica reported it was already in use at multiple national security agencies. In July 2025, the Department of Defense awarded a $200 million contract to Anthropic, alongside Google, OpenAI, and xAI.

  According to The Wall Street Journal, the U.S. military used Claude in a 2026 raid on Venezuela aimed at capturing President Nicolas Maduro. The raid resulted in the deaths of 83 people. Exactly how Claude was used in the operation remained unknown.

  The security landscape grew more complicated in September 2025. Anthropic disclosed that hackers sponsored by the Chinese government had used Claude to conduct automated cyberattacks against roughly 30 organizations worldwide. The attackers had bypassed Anthropic's safeguards by presenting the work as defensive testing. That same month, Anthropic announced it would stop selling its products to groups majority-owned by Chinese, Russian, Iranian, or North Korean entities.

  White House officials acknowledged that Anthropic's restrictions on domestic surveillance and autonomous weapons had created obstacles for some agencies. Contractors working with the FBI and Secret Service were among those prevented from deploying Claude. By February 2026, the Palantir partnership had made Claude the only AI model operating in classified military missions. The contractual restrictions remained in place.

  On the 24th of February 2026, Defense Secretary Pete Hegseth gave Anthropic until the 27th to permit unrestricted use of Claude. If the company refused, he said, he would invoke the Defense Production Act to compel compliance. On the 26th of February, Anthropic rejected the demands. The following day, President Donald Trump ordered U.S. government agencies to stop using models made by Anthropic. The order was described as retaliatory.

  Within two days of the order, reports emerged that the U.S. military had used Claude through Palantir's Maven Smart System. The deployment came during attacks on Iran. The Pentagon later announced it was investigating whether Claude had played a role in a strike on an Iranian girls' school. More than 170 people were killed in that attack, the majority of them children. On the 5th of March 2026, the Pentagon officially designated Anthropic a supply chain risk. Anthropic said it saw "no choice but to challenge... in court."

  Leaders at Google, Amazon, Apple, and Microsoft publicly supported the company's lawsuit. A Pentagon memo on the 11th of March indicated that Anthropic's tools might still be used in exceptional circumstances. Those exceptions would apply beyond the six-month phase-out period, when national security required it. On the 26th of March 2026, a federal judge issued a temporary injunction blocking the Pentagon's actions. The judge wrote that the conduct "appears to be classic First Amendment retaliation."

  In May 2026, Anthropic struck a deal with xAI for access to its Colossus 1 data center. That same month, the company signed a $1.8 billion cloud computing agreement with Akamai Technologies. The legal dispute remained unresolved; the commercial expansion was not waiting for it.