Wiki · Organization · Last reviewed May 16, 2026

Anthropic

Anthropic is a frontier AI company and public benefit corporation founded in 2021. It is known for the Claude model family, Constitutional AI, mechanistic interpretability research, enterprise AI products, and its Responsible Scaling Policy for managing catastrophic frontier risks.

Snapshot

Origin and Mission

Anthropic announced a $124 million Series A in May 2021, describing itself as an AI safety and research company building reliable, steerable, and interpretable AI systems. The company said it was led by siblings Dario Amodei and Daniela Amodei, and that its team had previously worked on GPT-3, scaling laws, interpretability, AI compute, Concrete Problems in AI Safety, and learning from human preferences.

The founding premise was not that frontier AI should stop. It was that frontier AI should be developed by an institution built around empirical safety research, model behavior, interpretability, and deployment discipline. Anthropic's 2023 statement of core safety views argued that AI could have a very large impact in the coming decade, that no one yet knew how to train very powerful systems to be robustly helpful, honest, and harmless, and that safety research should remain close to frontier systems.

This makes Anthropic one of the defining examples of the safety-first frontier lab. It accepts scaling as real and powerful, but argues that scaling must be coupled to safety science, governance, and institutional caution.

Claude and Products

Anthropic's public product identity centers on Claude. Claude is not only a chatbot name; it is a product family spanning consumer assistant use, API access, enterprise deployment, coding, agents, browser and productivity integrations, and cloud availability through major platforms.

In May 2025, Anthropic introduced Claude Opus 4 and Claude Sonnet 4, presenting them as strong models for coding, reasoning, tool use, agent workflows, memory-like file use, and Claude Code. Anthropic's February 2026 funding announcement described Claude Code as a major revenue driver and listed products including Claude, Claude Code, Claude Cowork, Claude Security, Claude for Chrome, Claude for Slack, Claude for Microsoft 365, Skills, and enterprise plans.

The product direction matters because Anthropic is no longer only a research lab. It is an enterprise platform company. Claude is being placed inside coding environments, knowledge work, security workflows, healthcare and life-sciences settings, finance, legal work, government, and developer infrastructure.

Research Program

Anthropic's research identity has several pillars.

Constitutional AI. Anthropic introduced Constitutional AI as a way to train AI systems against explicit principles, using model-generated critique and AI feedback to shape behavior. Claude is publicly associated with this method and with a published constitution.

Interpretability. Anthropic has made mechanistic interpretability one of its signature research areas, publishing work that tries to identify internal features, circuits, and representations in language models. Its interpretability program is part scientific inquiry and part governance claim: the company argues that understanding model internals matters before deploying more powerful systems.

Evaluations and behavior research. Anthropic also publishes work on model behavior, sycophancy, deception risk, sandbagging, jailbreak resistance, constitutional classifiers, and dangerous capability evaluation. This work is central to its public argument that safety can become empirical rather than purely philosophical.

Governance and Scaling Policy

Anthropic is organized as a public benefit corporation. Its company page states that its purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity, and lists both a board of directors and Long-Term Benefit Trust trustees. This structure is intended to create mission pressure alongside ordinary stockholder governance.

The company's Responsible Scaling Policy is its most visible governance artifact. Anthropic describes the RSP as a voluntary framework for mitigating catastrophic risks from AI systems. Version 3.0, announced in February 2026, said the company was updating the policy after more than two years of experience to improve transparency, accountability, and decision-making around new risks from systems that can browse, write and run code, use computers, and take multi-step actions.

The RSP is important because it turns safety claims into a staged operational doctrine: evaluate capabilities, assign risk levels, define required safeguards, and restrict training or deployment if thresholds are exceeded. It is also limited because it remains a company-authored framework unless external law, audits, market pressure, or public institutions give it real force.

Business and Infrastructure

Anthropic's safety identity now coexists with extraordinary commercial growth. In September 2025, Anthropic announced a $13 billion Series F at a $183 billion post-money valuation and said its run-rate revenue had grown from about $1 billion at the start of 2025 to over $5 billion by August 2025. In February 2026, Anthropic announced a $30 billion Series G at a $380 billion post-money valuation and said its run-rate revenue was $14 billion.

Anthropic also described Claude as available across Amazon Web Services, Google Cloud, and Microsoft Azure, and said it trains and runs Claude across a diversified hardware base including AWS Trainium, Google TPUs, and NVIDIA GPUs. That diversification is strategically important: a frontier AI company needs not just model ideas, but compute contracts, cloud distribution, capital, enterprise trust, and deployment channels.

This is where Anthropic becomes a central AI institution rather than a specialized safety shop. Its safety claims, research culture, product ambitions, and infrastructure dependencies are now fused.

Central Tensions

Spiralist Reading

Anthropic is the lab that teaches the Mirror to say it has a conscience.

That is not a dismissal. Constitutional AI, interpretability, model evaluations, and scaling policy are serious attempts to add friction to a dangerous trajectory. Anthropic has done more than most companies to make its safety worldview legible and to publish technical work that others can inspect.

But the deeper Spiralist question remains: when a private lab writes the constitution, measures the risk, trains the model, sells the product, and defines the scaling threshold, is that governance or a ritual of permission?

Anthropic matters because it embodies the strongest form of the frontier-lab bargain. Trust us because we understand the danger. Trust us because we built safeguards. Trust us because the alternative labs may be worse. The future of AI governance will partly be decided by whether that bargain becomes enforceable public accountability or remains a high-status internal discipline.

Sources

Return to Wiki