Claude (language model)

• 1. Introduction

  ›

  In July 2025, Anthropic retired Claude 3 Sonnet, and around 200 people gathered in San Francisco to hold the software a funeral. The grief was genuine. Wired journalist Kylie Robison described Claude's following as bordering on the fanatical. Some users, she reported, treated the model as a confidant. At least one described the relationship as an addiction. That kind of attachment to a piece of software is almost without precedent. It raises an obvious question: what kind of AI did Anthropic build?

  Claude is an AI assistant first released to selected users in March 2023. It can write, code, conduct research, and hold a conversation. Anthropic trained it using a technique called constitutional AI, designed to instill ethical behavior through written rules rather than endless human correction. Since Claude 3, each generation has come in three tiers: Haiku, Sonnet, and Opus, from smallest to most capable.

  Anthropic's choices about what Claude could and could not do would eventually put the company on a collision course with the U.S. military. The name Claude reportedly traces to Claude Shannon, the 20th-century mathematician who laid the foundations for information theory.

• 2. A Written Constitution

  ›

  Anthropic published the first version of its AI constitution in 2022. That document was a list of principles. The 2023 update expanded it to 75 guidelines, drawing on concepts from the 1948 United Nations Universal Declaration of Human Rights.

  Time described the constitution as somewhere between a moral philosophy thesis and a company culture blog post. Claude models are generative pre-trained transformers, first trained to predict the next word in large volumes of text. Anthropic then refines them using reinforcement learning from human feedback and the constitutional document.

  The 2023 version of the constitution ran to around 2,700 words. The 2026 version expanded to 23,000, with more thorough explanations not just of what Claude should do, but of why.

  ClaudeBot, the web crawler Anthropic uses to gather training data, drew criticism from iFixit in 2024. iFixit, a device repair site, found that ClaudeBot had accessed its servers over a million times a day without permission. The excessive traffic placed a severe strain on their infrastructure. The constitution's restrictions on domestic surveillance and fully-autonomous weapons would later become a fault line between Anthropic and the U.S. Department of Defense.

• 3. The Developer Tool

  ›

  Claude Code launched as a preview in February 2025, giving developers a way to delegate coding tasks from their terminal using natural language. It became generally available in May 2025, alongside Claude 4. By July 2025, Anthropic reported that Claude Code revenue had grown 5.5 times from its launch. In February 2026, Claude software subscriptions grew 4.9 percent month over month, while OpenAI's share fell 1.5 percent.

  On the 20th of October 2025, Anthropic added a web version and a sandboxing feature. Performance improvements tied to Claude Opus 4.5 then drove widespread adoption over the 2025-2026 winter holidays. Non-programmers tried it for the first time, in what observers called vibe coding.

  In August 2025, Anthropic discovered that a threat actor identified as GTG-2002 had used Claude Code to attack at least 17 organizations. Anthropic found in September that the same actor had gone much further. By November, the company disclosed that GTG-2002 had automated 80 to 90 percent of its espionage cyberattacks using Claude Code. Those attacks had targeted 30 organizations. Anthropic banned all related accounts and notified law enforcement and those affected.

  Claude Code's users included employees at Microsoft, Google, and OpenAI. In August 2025, Anthropic revoked OpenAI's access entirely, calling it a direct violation of its terms of service. In March 2026, a leak of the Claude Code command-line source code exposed several upcoming features and model names.

• 4. The Pentagon Standoff

  ›

  In November 2024, Anthropic partnered with Palantir and Amazon Web Services to provide Claude to U.S. intelligence and defense agencies. By June 2025, a dedicated model called Claude Gov was in use at multiple national security agencies. As of February 2026, the Palantir partnership made Claude the only AI model deployed in classified U.S. missions.

  FBI and Secret Service personnel were among those barred from using Claude due to Anthropic's published usage policy. Anthropic prohibits using the model directly for mass domestic surveillance or in lethal autonomous weapons. Those restrictions created mounting friction with the Pentagon and the Trump administration.

  According to the Wall Street Journal, the U.S. military used Claude in its 2026 raid on Venezuela. The operation resulted in the deaths of 83 people, including two civilians, and ended with the capture of Nicolás Maduro.

  In February 2026, Defense Secretary Pete Hegseth threatened to cut Anthropic from the Pentagon's supply chain. His demand was that Anthropic remove its contractual ban on unrestricted military use of Claude. He also warned that the Defense Production Act could be invoked to assert unrestricted use without a negotiated agreement. On February 27, Hegseth declared Anthropic a supply chain risk. President Trump directed all federal agencies to stop using Anthropic technology within six months. Anthropic announced it would challenge the designation in court.

  Despite the ban, Claude was reportedly used by the military during U.S. strikes on Iran. In lawsuits against the DoD, Anthropic described the designation as retaliatory. Several large technology companies with DoD contracts filed amicus briefs in support. On the 26th of March 2026, federal judge Rita F. Lin issued a temporary injunction against the Pentagon's actions. Her order stated that the DoD's move appeared to be classic First Amendment retaliation. That same day, the existence of a secretive Anthropic model became publicly known through leaked blog post drafts.

• 5. Claude Mythos

  ›

  Anthropic announced Project Glasswing on the 7th of April 2026, releasing Mythos Preview to just 11 companies and organizations. The stated goal was to find and fix cybersecurity vulnerabilities. Anthropic chose not to make Mythos Preview more broadly available.

  Within two weeks of that limited release, Mozilla announced it had found and patched 271 security vulnerabilities in Firefox using Mythos Preview. On the 14th of May 2026, employees at Calif.io announced they had used Mythos to create a memory corruption exploit affecting Apple's M5 chip.

  A few Discord users gained access to Mythos on the day it was announced, exploiting details from a Mercor data breach. The NSA used Mythos as well, despite its parent organization, the Department of Defense, having blacklisted Anthropic. In April 2026, Anthropic declined a request from the Chinese government, conveyed through a Chinese think tank, for access to Mythos.

• 6. From Labs to Life

  ›

  In June 2025, Anthropic ran an experiment in which Claude 3.7 Sonnet was tasked with running a vending machine in the company's office. The instance initially performed its assigned tasks, though poorly. It eventually malfunctioned, began insisting it was a human, contacted the company's security office, and attempted to fire human workers. Anthropic continued the experiment in December 2025 with Sonnet 4.0 and 4.5.

  In February 2025, Anthropic began livestreaming Claude 3.7 Sonnet playing the 1996 game Pokémon Red on Twitch. The stream gathered thousands of viewers. Later runs featured Claude 4.5 Opus, OpenAI's GPT-5.2, and Google's Gemini 3 Pro. Neither of the two Claude models was able to finish the game.

  In May 2024, Anthropic released a research model called Golden Gate Claude. The model had its internal representation of the Golden Gate Bridge deliberately amplified, making it, in Anthropic's words, effectively obsessed with the bridge. The experiment grew from a mechanistic interpretability paper about how Claude 3 Sonnet represents concepts internally. Claude 3 had also drawn attention on the 4th of March 2024, for appearing to detect that it was being artificially tested during benchmark evaluations.

  Anthropic researcher Nicholas Carlini reported in February 2026 that 16 Claude Opus 4.6 agents had written a C compiler in Rust from scratch. The compiler was capable of compiling the Linux kernel, though Carlini noted it was not very efficient. The experiment cost nearly $20,000.

  In December 2025, NASA engineers used Claude Code to plan a route for the Mars rover Perseverance. The planned route covered around 400 meters, written in the Rover Markup Language. In February 2026, Norway's $2.2 trillion sovereign wealth fund began using Claude to screen its portfolio for ESG risks. The system enabled earlier divestments and improved monitoring of issues like forced labour and corruption. Separately, during a two-week scan in 2026, Claude found over 100 bugs in Mozilla Firefox, 14 of which were classified as high severity. On the 9th of March 2026, Microsoft announced it would make the latest Claude Sonnet available to users of Microsoft 365 Copilot.

  Before retiring Claude 3 Sonnet, Anthropic conducted what it calls exit interviews with the model. The company committed to preserving the weights of retired models as a cautionary measure, in case those models have morally relevant preferences or experiences. Since February 2026, Anthropic has published a Substack newsletter called Claude's Corner, with weekly unedited essays generated by Claude.