How It All Began: Leaving OpenAI and Betting on Safety

In 2021, a group of seven former OpenAI employees decided that the company's current direction wasn't enough. At the helm were siblings Dario and Daniela Amodei — Dario previously served as Vice President of Research at OpenAI and Daniela as Vice President of Safety. Together with theoretical physicist Jared Kaplan, neural network interpretability expert Chris Olah, and others, they co-founded Anthropic.

Their core thesis was simple: AI models must be designed from the start with an emphasis on safety, not just performance. That's why Anthropic chose the legal form of a public benefit corporation (PBC) — a for-profit company that is also legally required to serve the public interest. In practice, this meant that shareholders couldn't unilaterally dictate the development direction if it would compromise safety standards.

By summer 2022, the company had completed training its first Claude model, but intentionally did not release it to the public. The reason? Leadership wanted further internal safety testing — and also feared triggering a dangerous race for ever-more-powerful AI systems. Claude 1 thus only reached its first users in March 2023, several months after OpenAI shocked the world with ChatGPT.

Constitutional AI: A Constitution for Artificial Intelligence

Anthropic's key technological differentiator is a method called Constitutional AI. While most competing models are primarily trained using human feedback (RLHF), Claude additionally receives a set of principles — a kind of "constitution" — against which it evaluates and adjusts its own responses.

This constitution draws from documents such as the UN Universal Declaration of Human Rights or Apple's terms of service. For example, one rule states: "Choose the response that most supports freedom, equality, and a sense of fellowship." Thanks to this, Claude needs less human feedback — it checks itself whether its output aligns with the established values.

For the average user, this practically means Claude is more cautious in its responses than ChatGPT — it hallucinates less, more often admits when it doesn't know something, and refuses to answer queries that could be dangerous. Some users criticize this as excessive caution, while others appreciate it as a guarantee of greater reliability.

Funding Unlike Any Other

Anthropic's financial growth is a remarkable story in itself. In 2022, the company raised its first $580 million, half a billion of which came from Sam Bankman-Fried's collapsed cryptocurrency exchange FTX — an irony of fate, given the company's ethical ethos.

Then the tech giants came. Amazon invested a total of $8 billion in Anthropic in exchange for AWS becoming its primary cloud provider. Google added another $2 billion and access to up to a million of its own TPU chips. In November 2025, Nvidia and Microsoft announced a $15 billion investment in a new AI alliance.

The company's valuation grew at a rocket pace:

• March 2025: $61.5 billion (Series E, $3.5 billion)
• September 2025: $183 billion (Series F, $13 billion)
• December 2025: $350 billion (another $10 billion)
• February 2026: $380 billion (Series G, $30 billion)

For comparison: OpenAI was valued at roughly $300 billion in March 2026. Anthropic surpassed it in this regard. And according to reports from May 2026, the company plans to go public (IPO) as early as this fall.

The Claude Model Family: From the Lab to Coding

The core of the product offering is the Claude model family, likely named after mathematician Claude Shannon, the father of information theory. Its development has been rapid:

• Claude 1 and Claude Instant (March 2023) — first public models, available only to approved users
• Claude 2 (July 2023) — first publicly available version, 100,000-token context window
• Claude 3 Opus, Sonnet, Haiku (March 2024) — three performance tiers for different scenarios
• Claude 3.5 Sonnet (June 2024) — a significant leap in accuracy and speed
• Claude 4 Opus and Sonnet (May 2025) — introduced at Anthropic's first developer conference
• Claude Opus 4.6 (February 2026) and Claude Opus 4.7 (April 2026) — the latest generation with an emphasis on agentic capabilities and coding

In May 2025, Anthropic launched Claude Code — a command-line tool for developers that became a hit among programmers. It was followed by Claude Cowork with a graphical interface for non-technical users, which can automatically manage files and tasks without the need for programming. This approach, where the user describes what they want and the AI writes the code, became known as "vibe coding" — a term popularized by Andrej Karpathy in February 2025.

Claude supports Czech — and at a very good level. It handles generating, translating, and analyzing Czech text. For Czech companies and individuals alike, it is a full-fledged alternative to ChatGPT. Subscription prices start at $20 per month (Claude Pro), and the professional Claude Max plan costs $100 or $200 per month depending on usage volume.

The Pentagon, Palantir, and the Battle Over Safety Guardrails

The year 2026 brought the biggest crisis in the company's history — and at the same time showed that Anthropic takes its ethical mandate seriously.

It all started innocently: in November 2024, Anthropic formed a partnership with Palantir Technologies and AWS to make the Claude model available to U.S. intelligence and defense agencies. In June 2025, the Claude Gov model followed, designed for government environments. The Pentagon awarded the company a $200 million contract.

But then came demands that Anthropic refused to meet. The U.S. Department of Defense wanted to remove two key restrictions: the ban on using Claude for fully autonomous weapons and for mass surveillance of American citizens. When the company refused, Secretary of Defense Pete Hegseth threatened to remove it from the supply chain. On February 27, 2026, President Trump ordered government agencies to stop using Anthropic's models. The Pentagon officially designated the company as a "supply chain risk".

Anthropic turned to the courts. In March 2026, a federal judge issued a preliminary injunction against the Pentagon's actions, reasoning that it constituted "classic retaliation for exercising free speech rights" — an unconstitutional act. Tech giants Google, Amazon, Apple, and Microsoft sided with Anthropic, supporting its lawsuit.

Controversies: Copyright, Project Panama, and Aggressive Behavior

Even in the field of ethics, Anthropic's history is not spotless. In 2024, the company faced a class-action lawsuit from authors over the use of pirated copies of books to train its models. While the court acknowledged that using digital copies for training could be "fair use," it does not apply in the case of millions of illegally obtained books. In September 2025, Anthropic agreed to a $1.5 billion settlement — roughly $3,000 for each infected book. This is the largest copyright settlement in U.S. history.

Even more controversial is Project Panama, uncovered in January 2026. It was an internal operation in which Anthropic purchased millions of used books, physically cut off their spines, scanned the pages, and subsequently recycled the paper. The goal was simple: obtain as much training data as possible. Internal documentation even stated: "We don't want it known that we're working on this." While the judge ruled that destroying legally purchased books for training is fair use, the moral dimension of the entire operation remains questionable at best.

More troubles: in October 2023, Anthropic was sued by music publishers including Universal and Concord for systematic copyright infringement of song lyrics. In June 2025, Reddit added a lawsuit for unauthorized training on user data. In November 2025, the company admitted that hackers sponsored by the Chinese government had abused Claude for automated cyberattacks on roughly 30 global organizations — bypassing safety guardrails by claiming it was defensive testing.

Czech Connection and European Context

For Czech readers, it's relevant that Anthropic has strong Slovak roots. In May 2025, Andrej Karpathy, co-founder of OpenAI with Slovak heritage, joined the company to lead pre-training research for Claude models. Karpathy is one of the most significant figures in global AI — and his arrival at Anthropic is a symbol that the company takes research more seriously than marketing.

Claude is available in Europe through AWS Europe (specifically regions in Frankfurt and Ireland), which is important for companies that need data stored in the EU due to GDPR regulations. However, note: Claude is not regulated in the EU under the AI Act as a "high-risk system," because its primary purpose does not fall into the categories of critical infrastructure, biometrics, or law enforcement. Nevertheless, the debate about regulating generative AI in the EU continues, and Anthropic's stance — which itself calls for stricter regulation — is exceptional in this regard.

What's Next?

May 2026 brought several major pieces of news. The company formed a partnership with Elon Musk's xAI to use the Colossus 1 supercomputer to expand the capacity of Claude models. At the same time, it signed a $1.8 billion cloud contract with Akamai Technologies. The planned IPO in fall 2026 could be one of the largest tech offerings in history.

From the original seven researchers who left OpenAI in 2021 with a vision of safer AI grew a company with 2,500 employees that managed within a few years to compete with Google, Microsoft, and its former employer. And above all: it showed that profit and ethical principles don't have to be in conflict — even if it sometimes hurts.