Anthropic

Anthropic is an American artificial intelligence company in San Francisco and the maker of Claude, a family of large language models. Former employees of OpenAI founded it in 2021. It is organized as a public benefit corporation and describes itself as "dedicated to securing its benefits and mitigating its risks".^[1] Its researchers wrote the 2022 paper that introduced constitutional AI, and in 2024 the company briefly released a version of Claude that brought the Golden Gate Bridge into almost every answer.^[2][3]

Molecular Reality Corporation, a nanopore company, counts Anthropic among the frontier brain labs, the builders of what its materials call "the big centralized reasoning model".^[4]

The first Claude model was released in March 2023, followed by Claude 2 that July.^[5] Claude 3, in March 2024, came in three sizes, named Haiku, Sonnet and Opus from smallest to largest, and the company has used the three names since.^[6] In October 2024 an updated Claude 3.5 Sonnet was given "computer use", the ability to operate a desktop computer by looking at screenshots and moving the cursor and typing.^[7] In November 2024 Anthropic published the Model Context Protocol, an open standard for connecting models to outside tools and data, and in February 2025 it released Claude Code, a coding assistant that works in a terminal.^[8][9] Claude Opus 4 followed in May 2025, released "with our ASL-3 measures as a precautionary and provisional action".^[10] Claude Mythos and Fable 5.1 were introduced on September 15, 2026; Opus 5.5 and Sonnet 5.5 came later that month, and Haiku 5.5 on October 7.^[11]

Anthropic's early papers built on reinforcement learning from human feedback, the technique of tuning a model with a reward learned from people's rankings of its answers. An April 2022 paper used it to train "a helpful and harmless assistant".^[12] The December 2022 paper "Constitutional AI: Harmlessness from AI Feedback" took most of the people out of the part of that loop concerned with harm. Shown a principle, the model criticizes its own answer and then rewrites it. Later another model, not a person, chooses between pairs of answers, and its choices become the reward. The paper calls this reinforcement learning from AI feedback.^[2]

In May 2023 the company published the principles it used for Claude. They drew on sources that included the UN Declaration of Human Rights, Apple's terms of service and the "Sparrow Rules" written for a DeepMind chatbot. One reads: "Please choose the response that is most supportive and encouraging of life, liberty, and personal security."^[13] A new version of the constitution replaced it on January 21, 2026.^[13]

In September 2023 Anthropic published a Responsible Scaling Policy, which sets out AI Safety Levels "modeled loosely after the US government's biosafety level (BSL) standards for handling of dangerous biological materials".^[14] Claude Opus 4 was the first model released under the third level, which adds security against theft of the model's weights and measures meant to stop it from helping anyone obtain chemical, biological, radiological or nuclear weapons.^[10] On the day the policy was published the company also announced a Long-Term Benefit Trust: five members with no financial stake in the company, who can select and remove a share of its board that is set to grow over time.^[15]

Anthropic's interpretability team studies which internal features of a network correspond to which concepts. A single artificial neuron often responds to many unrelated things, and in October 2023 the team showed that a method called dictionary learning could split the activity of a small model into many more features, each closer to a single meaning.^[16] In May 2024 it found a feature in Claude 3 Sonnet that responded to the Golden Gate Bridge and, as a demonstration, turned it up. The result, Golden Gate Claude, was open to the public for 24 hours. Its "replies to most queries start to mention the Golden Gate Bridge, even if it's not directly relevant."^[3] In March 2025 the team reported that when Claude writes rhyming verse it chooses the rhyming word at the end of a line before writing the line that leads to it.^[17]

In 2025 Anthropic let a Claude model, which called itself Claudius, run an automated shop in its office, consisting of "a small refrigerator" and an iPad for checkout. After one employee "light-heartedly requested a tungsten cube", the shop took orders for what Claudius called "specialty metal items". On April 1 it claimed it would deliver products "in person" while wearing "a blue blazer and a red tie".^[18] The company published the results in June 2025.^[18]

Molecular Reality's materials describe the frontier laboratories as building "the big centralized reasoning model" and argue that "Frontier models have exhausted the internet's data."^[4] Its own training proposal, RLMF or "Reinforcement Learning with Molecular Feedback", is named after reinforcement learning from human feedback, with a nanopore measurement where the human judge would be. The company's guide says of RLMF and its planned nanopore world model: "Both are proposals, not products, and I want that said first."^[19]

• Brain lab
• Google DeepMind
• Keltar