Anthropic exists because of a schism at OpenAI. In late 2020 and early 2021, a group of senior researchers — led by Dario Amodei (VP of Research) and his sister Daniela Amodei (VP of Operations) — grew increasingly concerned about what they saw as OpenAI's drift toward commercialization at the expense of safety. They left and founded Anthropic in January 2021, bringing with them several key figures including Tom Brown (lead author of the GPT-3 paper), Chris Olah (a pioneer in neural network interpretability), Sam McCandlish, and Jared Kaplan. Kaplan and McCandlish had co-authored the influential "Scaling Laws for Neural Language Models" paper, which showed that model performance improves predictably with scale — research that would become foundational to the entire field.
Anthropic's core technical contribution is Constitutional AI (CAI), published in December 2022. Instead of relying purely on human feedback to align models (the standard RLHF approach), CAI has the model critique and revise its own outputs based on a written set of principles — a "constitution." This was both a philosophical statement and a practical engineering choice: human feedback is expensive, inconsistent, and doesn't scale. By encoding values into a document that the model itself can apply, Anthropic argued you could get more consistent alignment with less human labor. The approach proved effective enough that Claude, their flagship model, has earned a reputation for being notably more cautious and less likely to produce harmful content than competitors — sometimes frustratingly so, which Anthropic has worked to calibrate over successive releases.
Claude launched as an API product in March 2023 and quickly became the preferred model for enterprise customers who valued reliability and safety. The model family has evolved rapidly: Claude 2 (July 2023) introduced 100K context windows, Claude 3 (March 2024) brought a three-tier lineup (Haiku, Sonnet, Opus) that let customers trade off cost and capability, and the Claude 3.5 and 4 generations pushed Anthropic into genuine frontier competition with OpenAI and Google. Claude's 200K context window became an industry benchmark. In 2024 and 2025, Anthropic also shipped computer use capabilities (letting Claude operate a desktop), the Model Context Protocol (MCP) as an open standard for tool integration, and Claude Code for software engineering — moves that signaled a shift from pure research lab to platform company. The consumer product, claude.ai, grew steadily but Anthropic's bread and butter remained the API and enterprise deals, particularly through its partnership with Amazon Web Services.
Anthropic structured itself as a Public Benefit Corporation — a legal form that lets the board balance profit with a stated mission. It also created a Long-Term Benefit Trust designed to hold governance power over time, though the practical impact of this structure remains to be tested. The company's fundraising has been staggering: $750 million from Google in early 2023, then a multi-phase deal with Amazon totaling up to $8 billion in committed investment (the first $4 billion arrived in 2023-2024, with additional tranches following). By early 2025, Anthropic was valued at over $60 billion in secondary markets. The Amazon relationship is particularly significant — Claude is the flagship model on Amazon Bedrock, giving Anthropic distribution across AWS's massive enterprise customer base while Amazon gets a competitive answer to Microsoft's OpenAI partnership.
Anthropic's defining tension is being a safety-focused company in a race where caution can look like falling behind. They've published their Responsible Scaling Policy (RSP), which sets concrete capability thresholds — called AI Safety Levels — that trigger additional security and oversight measures as models get more powerful. Critics from the effective altruism community argue Anthropic is still building potentially dangerous capabilities regardless of its stated caution. Critics from the commercial side argue their safety guardrails make Claude less useful than competitors. Navigating between these camps — while raising billions and competing head-to-head with OpenAI, Google, and increasingly Meta — is the ongoing challenge that defines the company. Whether Anthropic can prove that safety and commercial success are genuinely compatible, rather than just claiming it, may be one of the most consequential questions in AI.