Anthropic
Anthropic is a frontier AI safety company founded in 2021 by former OpenAI VP of Research Dario Amodei and his sister Daniela Amodei. After leaving OpenAI, they publicly worried that it was "trading safety for commercialization," so Anthropic put AI safety first — every research, product, and business decision has to clear a safety bar.
Key milestones
- 2021 founded as an AI safety company
- 2023-03 Claude 1 released, conversation quality made the cut
- 2023-07 Claude 2 with long context + Constitutional AI
- 2024-03 Claude 3 family: Opus / Sonnet / Haiku tiers
- 2024-06 Claude 3.5 Sonnet leapfrogged on coding
- 2024-11 Computer Use (model can drive a desktop / browser)
- 2025-01 Claude 3.7 Sonnet with hybrid reasoning
- 2025-02 Claude Code launched as an agentic dev tool
- 2024-11 open-sourced MCP; within half a year it became the de-facto standard for agent tool calls
Flagship products
- Claude family: see
claude.md - Claude Code: terminal-native agent that drives Claude from the command line for development tasks
- Computer Use: model directly operates browsers and desktops
- MCP: open protocol for LLM access to external data / tools
Key methods
- Constitutional AI (CAI): a set of "constitutional principles" lets the model self-critique and self-revise, replacing the human feedback loop with an AI feedback loop (RLAIF). Anthropic's signature approach.
- Mechanistic interpretability: long-term investment in understanding transformer internals — "opening the black box."
- Safety evaluation at scale: every new training run comes paired with red-team testing and capability evaluation.
Influence
- Claude set long-standing benchmarks in long context, code, agents, and document processing
- MCP was rapidly adopted by OpenAI, Google, Cloudflare, Cursor, Replit — Anthropic used open source to win the agent tool-call standard
- Proved the narrative "AI safety ≠ slower release": Claude 3 arrived 2 years after GPT-4 but earned strong market acceptance
- Valuation passed $180B in 2025, second only to OpenAI among frontier AI labs
Limitations
- Limited public detail on training data and compute
- Deeply tied to AWS (major cloud + investor), multi-cloud strategy is constrained
- Commercialization lags OpenAI by a beat — Claude Code / Computer Use came after OpenAI's analogous products