Zum Inhalt springen

Anthropic and Claude

Zusammenfassung

Anthropic is the AI lab built by people who walked out of the room. In 2020 a group of senior OpenAI researchers — led by the sibling pair Dario and Daniela Amodei — concluded that the field was scaling faster than it was learning to control what it scaled, and left to build a lab where safety was the founding premise rather than a side team. Founded in January 2021 as a public benefit corporation, Anthropic invented Constitutional AI (training a model to critique itself against a written set of principles), shipped the Claude family of assistants, open-sourced the Model Context Protocol that became the industry’s plug standard for tool use, and turned Claude Code into one of the fastest-growing developer tools in history. By 2026 it had taken tens of billions from Amazon and Google, reached a valuation in the hundreds of billions, and filed to go public — all while its CEO kept publishing essays warning that the technology his company sells could go catastrophically wrong.

The Walkout

Anthropic’s origin is a schism. In 2020, Dario Amodei — OpenAI’s VP of Research and a lead author of the “scaling laws” papers showing that model capability rises predictably with compute and data — grew convinced that OpenAI’s commercial turn (the 2019 Microsoft deal, the capped-profit restructuring) was outrunning its safety commitments. He and his sister Daniela Amodei, who ran OpenAI’s safety and policy operations, left at the end of 2020, taking a cohort of senior researchers with them.

The founding team that incorporated Anthropic in January 2021 included Dario (CEO) and Daniela (President) plus Jared Kaplan (the physicist behind the scaling laws), Chris Olah (a pioneer of mechanistic interpretability — reverse-engineering what individual neurons in a network actually compute), Sam McCandlish, Tom Brown (lead engineer on GPT-3), and others. The bet was contrarian: that the path to safe AI ran through building frontier models, because you cannot align a system you cannot study.

Public Benefit Corporation

Anthropic chose to incorporate as a public benefit corporation (PBC) — legally obligated to weigh public good alongside shareholder return — and later added a Long-Term Benefit Trust, a body holding a special class of shares with the power to elect board members, intended to keep mission ahead of investors as the money got large. Whether such governance can actually constrain a company raising tens of billions remains the open question of the AI-safety era; it is the same structural problem that produced the OpenAI board crisis of November 2023, here attempted with a different legal scaffold.

Constitutional AI

Anthropic’s signature research contribution arrived in December 2022: Constitutional AI (CAI). The standard alignment method of the day, RLHF (reinforcement learning from human feedback), required armies of human labelers to rank model outputs — expensive, inconsistent, and psychologically taxing. CAI replaced much of that with a written “constitution”: a list of principles (drawn from sources like the UN Declaration of Human Rights) that the model uses to critique and revise its own answers. The model generates a response, judges it against the constitution, rewrites it, and trains on the improved version — “harmlessness from AI feedback.”

The technique made the values steering the model explicit and auditable rather than buried in labeler intuition, and it scaled supervision in a way pure human feedback could not. Every Claude model since has been shaped by it.

The Claude Family

Anthropic finished training its first model, Claude 1, in 2022 — before ChatGPT shipped — but, true to character, held it back rather than risk accelerating a race it feared. Claude reached the public via the Anthropic API on March 14, 2023 (Claude and the cheaper Claude Instant).

  • Claude 2 (2023) added a 100K-token context window, later pushed to 200K with Claude 2.1 — long-context understanding became a Claude signature.
  • Claude 3 (March 2024) introduced the durable three-tier naming — Opus (most capable), Sonnet (balanced), Haiku (fast/cheap) — and added vision.
  • Claude 3.5 Sonnet (mid-2024) and the Claude 4 generation (2025: Opus 4, Sonnet 4.5, Haiku 4.5) pushed hard into agentic work — long-horizon, multi-step tasks — making Claude the developers’ default for coding.
  • The Opus 4.x line continued through 2026 (Opus 4.8, May 2026), with the frontier extending into a new top tier branded Fable 5.

The naming — Opus, Sonnet, Haiku — quietly signals the house style: a model marketed less as an oracle than as a careful writer.

MCP and Claude Code

Two 2024–25 releases turned Claude from a chatbot into infrastructure:

  • The Model Context Protocol (MCP), open-sourced in November 2024, is an open standard for connecting AI assistants to external tools and data sources — a “USB-C port for AI.” Anthropic gave it away, and within a year competitors and tool-makers across the industry adopted it, making MCP one of the rare cases where a single lab’s protocol became a de-facto standard.
  • Claude Code (2025), a terminal-based agentic coding tool, let Claude read, edit, run, and commit code directly in a developer’s repository. It became one of the fastest-revenue-ramping developer products on record and a major driver of Anthropic’s enterprise business — coding emerging as the first domain where AI agents delivered clear, paid-for economic value.

The Money

Anthropic’s funding tracks the scale of the compute it consumes. After a $124M Series A in 2021, the company drew strategic investment from Google and then a landmark commitment from Amazon (Claude became a flagship model on AWS Bedrock; Anthropic adopted Amazon’s Trainium chips). By 2025 Amazon’s total commitment ran into the tens of billions, paired with a multi-year, $100B-plus AWS spend pledge from Anthropic; Google added its own multi-tens-of-billions stake. Valuation climbed from ~$61.5B (early 2025) to the hundreds of billions within roughly a year, and in 2026 Anthropic confidentially filed for an IPO — positioning one of the largest public offerings in tech history. Strikingly, the Anthropic stakes became a top profit driver in Amazon’s and Alphabet’s own earnings.

Dead End / Tension: The Safety Lab That Sells Frontier Models

Anthropic’s defining contradiction is structural, not accidental. Its founders left OpenAI warning that racing to build powerful AI is dangerous — and then built a company that races to build powerful AI, arguing it must be at the frontier to make the frontier safe. Dario Amodei’s 2024 essay “Machines of Loving Grace” paints AI’s upside in near-utopian terms (curing diseases, compressing a century of progress into a decade); his earlier and later writing warns of catastrophic and even existential risk. Anthropic publishes a Responsible Scaling Policy with capability thresholds (“AI Safety Levels”) meant to trigger stricter safeguards — but it also keeps shipping more capable models on a fast cadence, into the same competitive dynamics it once fled. Critics call this having it both ways; Anthropic calls it the only honest position available. The encyclopedia’s verdict is open: it is the central wager of the alignment era, and the result is not yet in.

Fun Fact: The Model That Sat in a Drawer

Anthropic had a working Claude-class model in 2022, before OpenAI released ChatGPT — and deliberately did not ship it, out of concern that being first would accelerate exactly the race dynamics the company was founded to slow. ChatGPT launched in November 2022 and triggered the race anyway. Claude reached the public four months later. The lab built to avoid lighting the fuse watched someone else light it, then joined the race it had hoped to forestall.

📚 Sources