Claude (language model)

Claude models are generative pre-trained transformers that have been trained to predict the next word in large amounts of text. Then, they have been fine-tuned using reinforcement learning from human feedback (RLHF) and constitutional AI in an attempt to enforce ethical guidelines. ClaudeBot searches the web for content. It respects a site's robots.txt but was criticized by iFixit in 2024, before they added their robots.txt, for placing excessive load on their site by scraping content. Constitutional AI Anthropic introduced an approach to AI alignment called "Constitutional AI". The constitution is a document used for training Claude to be harmless and helpful without relying on extensive or expensive human feedback. The original version was a list of principles, whereas the 2026 constitution explains more thoroughly how Claude is intended to behave and why. Anthropic said it intends Claude's constitution to be a model followed by others in the industry. The first constitution for Claude was published in 2022. The 2023 update listed 75 guidelines for Claude to follow. The first constitutions included concepts taken from the 1948 UN Universal Declaration of Human Rights. The philosopher Amanda Askell is the lead author of the 2026 constitution, with contributions from Joe Carlsmith, Chris Olah, Jared Kaplan, and Holden Karnofsky. The constitution is released under Creative Commons CC0. Time described this constitution as "somewhere between a moral philosophy thesis and a company culture blog post". In the supervised learning phase, the model generates responses to prompts, self-critiques these responses based on a set of guiding principles (a "constitution"), and revises the responses. Then the model is fine-tuned on these revised responses. For the reinforcement learning from AI feedback (RLAIF) phase, responses are generated, and an AI compares their compliance with the constitution. This dataset of AI feedback is used to train a preference model that evaluates responses based on how much they satisfy the constitution. Claude is then fine-tuned to align with this preference model. This technique is similar to RLHF, except that the comparisons used to train the preference model are AI-generated. == Features ==

Subscription plans Subscription plans offer access to additional and exclusive features such as Claude Code and Claude in Chrome. In May 2024, Anthropic began offering enterprise options with Claude Team and Claude Enterprise subscriptions for multiple users. These provide additional chats and other benefits. In April 2025, Anthropic released Claude Max, a higher-tier subscription offering more usage and access to early and exclusive features. The subscription comes with either a 5x option ($100 a month) or 20x option ($200 a month) with respective increases to the plan's usage limits. Web search In March 2025, Anthropic added a web search feature to Claude, starting with paying users in the United States. Free users gained access in May 2025. Artifacts In June 2024, Anthropic released the Artifacts feature, allowing users to generate and interact with code snippets and documents. Computer use In October 2024, Anthropic released the "computer use" feature, allowing Claude to attempt to navigate computers by interpreting screen content and simulating keyboard and mouse input. == Claude Code ==

Claude Code is an agentic command line tool released in February 2025 that enables developers to delegate coding tasks directly from their terminal using natural language prompts. While initially released for preview testing, it was made generally available in May 2025 alongside Claude 4. Based on enterprise adoption, Anthropic reported a 5.5x increase in Claude Code revenue by July. Anthropic released a web version that October as well as an iOS app. Coinciding with performance improvements tied to Claude Opus 4.5, Claude Code went viral during the 2025-2026 winter holidays when people had time to experiment with it, including many non-programmers who used it for vibe coding. In August 2025, Anthropic released Claude for Chrome, a Google Chrome extension allowing Claude Code to directly control the browser. In August 2025, Anthropic revealed that a threat actor called "GTG-2002" used Claude Code to attack at least 17 organizations. In November 2025, Anthropic announced that it had discovered in September that the same threat actor had used Claude Code to automate 80–90% of its espionage cyberattacks against 30 organizations. All accounts related to the attacks were banned, and Anthropic notified law enforcement and those affected. Google, and OpenAI employees. In August 2025, Anthropic revoked OpenAI's access to Claude, calling it "a direct violation of our terms of service". Claude Cowork is a tool similar to Claude Code but with a graphical user interface, aimed at non-technical users. It was released in January 2026 as a "research preview". Cowork is available as a desktop application for Mac and Windows and through a web interface. It provides Claude with access to a sandboxed shell and to user-selected folders on the local file system, allowing the model to read, write and edit files, execute code, and chain multi-step tasks within a single conversation, such as organising files on a user's computer or generating documents from material in a local folder. A Wall Street Journal reporter who tested the tool noted user concerns about granting an AI agent broad access to a personal computer. According to developers, Cowork was mostly built by Claude Code. In March 2026, the source code for the Claude Code command-line interface application was leaked, revealing multiple upcoming features and models. == Models ==

Claude is reportedly named after Claude Shannon, a 20th-century mathematician who laid the foundation for information theory. Claude models are usually released in three sizes: Haiku, Sonnet, and Opus (from smallest and cheapest to largest and the most expensive). Claude The first version of Claude was released in March 2023. Claude 2 Claude 2, released in July 2023, became the first Anthropic model available to the general public. Claude 3 Claude 3 was released on March 4, 2024. It drew attention for demonstrating an apparent ability to realize it is being artificially tested during 'needle in a haystack' tests. Claude 3.5 On June 20, 2024, Anthropic released Claude 3.5 Sonnet, which, according to the company's own benchmarks, performed better than the larger Claude 3 Opus. Released alongside 3.5 Sonnet was the new Artifacts capability in which Claude was able to create code in a separate window in the interface and preview in real time the rendered output, such as SVG graphics or websites. An upgraded version of Claude 3.5 Sonnet was introduced in October 22, 2024, along with Claude 3.5 Haiku. A feature, "computer use", was also released in public beta. This allowed Claude 3.5 Sonnet to interact with a computer's desktop environment by moving the cursor, clicking buttons, and typing text. This development allows the AI to attempt to perform multi-step tasks across different applications. Claude 4 On May 22, 2025, Anthropic released two more models: Claude Sonnet 4 and Claude Opus 4. Anthropic added API features for developers: a code execution tool, "connectors" to external tools using its Model Context Protocol, and Files API. It classified Opus 4 as a "Level 3" model on the company's four-point safety scale, meaning they consider it so powerful that it poses "significantly higher risk". Anthropic reported that during a safety test involving a fictional scenario, Claude and other frontier LLMs often send a blackmail email to an engineer in order to prevent their replacement. Claude Opus 4.1 (multi-factor authentication with TOTP and WebAuthn) In August 2025 Anthropic released Opus 4.1. It also enabled a capability for Opus 4 and 4.1 to end conversations that remain "persistently harmful or abusive" as a last resort after multiple refusals. Claude 4.5 Anthropic released Haiku 4.5 on October 15, 2025. Reporting by Inc. described Haiku 4.5 as targeting smaller companies that needed a faster and cheaper assistant, highlighting its availability on the Claude website and mobile app. Anthropic released Opus 4.5 on November 24, 2025. Claude 4.6 Anthropic released Opus 4.6 on February 5, 2026. The main improvements included an agent team and Claude in PowerPoint. , it is the model with the longest task-completion time horizon as estimated by METR, having a 50%-time horizon of 14 hours and 30 minutes and a 80%-time horizon of 1 hour 3 minutes. Anthropic released Sonnet 4.6 on February 17, 2026. Claude 4.7 Anthropic released Opus 4.7 on April 16, 2026. Claude Mythos The existence of a model named Claude Mythos had become publicly known on March 26 due to leaked blog post drafts. On April 7, 2026, Anthropic announced a preview of Claude Mythos ("Project Glasswing/Claude Mythos Preview") would be made available to 11 companies and organizations to find and fix cybersecurity vulnerabilities. Anthropic has stated that it does not plan to make Claude Mythos Preview generally available. Anthropic said that Mythos is capable of finding critical vulnerabilities in major operating systems and web browsers. The UK AI Security Institute reported that Claude Mythos was able to solve a 32-step cyber range called "The Last Ones" 3 out of 10 times, becoming the first model to do so end-to-end. (The next best models on this evaluation were Claude Opus 4.6, followed by a tie between GPT-5.4 and GPT-5.3 Codex. Reportedly, a few users in a private Discord channel gained access to Mythos the same day it was announced, using details from the recent Mercor data breach. The NSA has also used Mythos, despite the fact that the US Department of Defense, its parent organization, had blacklisted Anthropic after a dispute. == Model retirement==

Anthropic committed to preserve the weights of the retired models "for at least as long as the company exists"; the company also conducts "exit interviews" with models before their retirement. Anthropic gave its retired Claude 3 Opus model, deprecated in January 2026, its own Substack blog called "Claude's Corner". The newsletter will run for at least three months with weekly unedited essays. Claude 3 Opus was later brought back for paying customers and is available by request via API, though Anthropic still refers to it as "retired". == Research==

In May 2024, Anthropic issued a mechanistic interpretability paper identifying "features" (internal representations of concepts) in Claude 3 Sonnet, and released "Golden Gate Claude", a model for which the Golden Gate Bridge feature was strongly activated, leading Claude to be "effectively obsessed" with the bridge. In June 2025, Anthropic tested how Claude 3.7 Sonnet could run a vending machine in the company's office. The instance initially performed its assigned tasks, although poorly, until it eventually malfunctioned and insisted it was a human, contacted the company's security office, and attempted to fire human workers. In December 2025, the experiment continued with Sonnet 4.0 and 4.5. In November 2025, Anthropic tested Claude's ability to assist humans in programming a robot dog. In February 2025, Claude 3.7 Sonnet playing 1996 game Pokemon Red started to be livestreamed on Twitch, gathering thousands of viewers. Similar livestreams were later set with Claude 4.5 Opus, OpenAI's GPT-5.2, and Google's Gemini 3 Pro. Both Claude models were unable to finish the game. In February 2026, Anthropic's researcher Nicholas Carlini reported that 16 Claude Opus 4.6 agents were able to write a C compiler in Rust from scratch, "capable of compiling the Linux kernel". The experiment cost nearly $20,000; Carlini noted that even though the compiler is not very efficient, Opus 4.6 is the first model able to write it. == Usage ==

In December 2025, Claude was used to plan a route for NASA's Mars rover, Perseverance. NASA engineers used Claude Code to prepare a route of around 400 meters using 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, enabling earlier divestments and improved monitoring of issues like forced labour and corruption. During a two-week scan in 2026, Claude found over 100 bugs in the Mozilla Firefox web browser, of which 14 were considered high severity. On 9 March 2026, Microsoft said that it will be making the latest Claude Sonnet model available to Microsoft 365 Copilot users. Later that month, social media platform Bluesky released Attie, a Claude-based chatbot and curator. Military usage In November 2024, Anthropic partnered with Palantir and Amazon Web Services to provide the Claude model to U.S. intelligence and defense agencies. In June 2025, Anthropic announced a "Claude Gov" model. Ars Technica reported that as of June 2025 it was in use at multiple U.S. national security agencies. As of February 2026, Anthropic's partnership with Palantir makes Claude the only AI model used in classified missions. According to The Wall Street Journal, the U.S. military used Claude in its 2026 raid on Venezuela. While it isn't known to what capacity Claude was used, the operation resulted in the deaths of 83 people, two of which were civilians, and the capture of Nicolás Maduro. Anthropic's usage policy prohibits directly using Claude for domestic surveillance or in lethal autonomous weapons. These restrictions led to members of the FBI and Secret Service being unable to use it, and to tensions with the Pentagon and the Trump administration. In February 2026, the Financial Times reported that Defense Secretary Pete Hegseth threatened to cut Anthropic out of the DoD's supply chain if Anthropic did not permit unrestricted use of Claude, or to invoke the Defense Production Act to assert unrestricted use without an agreement. In lawsuits filed by Anthropic against the Department of Defense (DoD), Anthropic described the ban as retaliatory. Several large technology companies with DoD contracts filed amicus briefs in support of Anthropic. User base Wired journalist Kylie Robison wrote that Claude's "fan base is unique", comparing it to more ordinary ChatGPT users. In July 2025, when Anthropic retired its Claude 3 Sonnet model, around 200 people gathered in San Francisco for a "funeral". According to Robison, == See also ==