Anthropic is an AI safety and research company founded in 2021 by Dario Amodei, Daniela Amodei, and a group of colleagues who left OpenAI, where Dario Amodei had led research. It develops the Claude family of large language models and is distinguished among frontier laboratories by an explicitly safety-first public identity: the company frames its frontier-model development as a means of doing safety research on the systems that matter.1

Stated mission and record

Anthropic's company page describes it as "an AI safety and research company" that builds "reliable, interpretable, and steerable AI systems," and states that as a public-benefit corporation its purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."4 The 2023 essay "Core Views on AI Safety" supplies the rationale for building frontier models at all: "A major reason Anthropic exists as an organization is that we believe it's necessary to do safety research on 'frontier' AI systems."5

As of July 2026, the record shows both consistencies and tensions with this framing. Consistent with the mission, Anthropic published the field's first Responsible Scaling Policy with ASL capability gates, originated Constitutional AI, sustains a prominent Mechanistic interpretability program, and has issued unusually candid system cards, including the Claude 4 documentation that disclosed blackmail behavior in evaluation scenarios. At the same time, the company competes directly at the frontier it characterizes as potentially dangerous, shipping successive Claude generations against OpenAI and Google DeepMind through the Fable/Mythos 5 era, and pursuing large-scale commercialization despite a founding narrative tied to disagreements over commercialization at OpenAI. Supporters view frontier participation as the mission's stated method; critics contend it contributes to the racing dynamics the company itself warns about, and both readings remain current in public debate.

History

The founding team departed OpenAI in late 2020 amid what contemporaneous reporting described as differences over commercialization and safety direction. Anthropic organized as a public-benefit corporation, raised successive multi-billion-dollar rounds, including major investments from Google and Amazon (the latter tied to training on AWS silicon), and released its first Claude model publicly in March 2023, after ChatGPT had ignited the assistant market.1

The Claude line advanced through Claude 2 (2023), the Claude 3 generation (2024), the 3.5–4.x series (2024–2025), and into the Claude 5 generation in 2026; per Anthropic's announcements, the June 2026 Fable/Mythos 5 release marked its first "Mythos-class" tier above Opus.2

Research and models

Anthropic's research identity centers on alignment and interpretability: it originated constitutional approaches to post-training as an evolution of RLHF,3 and its interpretability group under Chris Olah drove much of the mechanistic-interpretability agenda, from circuits to sparse autoencoders. Its models are built on the Transformer stack and evaluated on the standard suites, including MMLU, where Claude 3 Opus was among the first to challenge GPT-4-class scores, as reported by the developer.12 Amodei has also argued the lab's optimistic case in public essays, most substantially the October 2024 Machines of Loving Grace, which sketches a "compressed 21st century" of AI-driven scientific progress, and the 2025 "The Urgency of Interpretability."6

Safety approach

Anthropic published the field's first Responsible Scaling Policy (2023), defining AI Safety Level (ASL) capability thresholds with corresponding security and deployment gates, a framework later echoed by OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework. It publishes system cards with major releases and collaborates with external evaluators on pre-deployment testing.1

See also

References


  1. Anthropic, company announcements and Responsible Scaling Policy, 2021–2025, anthropic.com; contemporaneous coverage of the founding in Reuters and The New York Times, 2021. ↩↩↩↩

  2. Anthropic, Claude release announcements, 2023–2026, including "Introducing Claude Sonnet 5" (June 2026). ↩↩

  3. Bai, Y., et al., "Constitutional AI: Harmlessness from AI Feedback," arXiv:2212.08073, December 2022. ↩

  4. Anthropic, "Company," official website, as of July 2026. ↩

  5. Anthropic, "Core Views on AI Safety: When, Why, What, and How," official website, March 2023. ↩

  6. Amodei, D., "Machines of Loving Grace," darioamodei.com, October 2024; Amodei, D., "The Urgency of Interpretability," darioamodei.com, 2025. ↩