Anthropic
2021
San Francisco, CA
AI safety, LLMs, constitutional AI, reasoning
Anthropic
Anthropic is the safety-first AI lab — the one that asks “how do we make this not go wrong?” as seriously as “how do we make this powerful?” Founded by people who left OpenAI specifically because they wanted a different approach to building frontier AI.
They built Claude (the model powering this very coding session), pioneered Constitutional AI (a way to align models using principles rather than just human ratings), and published the Responsible Scaling Policy — a commitment to test for dangerous capabilities before deploying more powerful models.
If OpenAI is “move fast,” Anthropic is “move carefully, but still move.”
The Models
| Model | Character |
|---|---|
| Claude 4 Opus | Deep reasoning, complex analysis, frontier capability |
| Claude 4 Sonnet | The sweet spot — fast, capable, excellent at coding |
| Claude 3.5 Haiku | Quick and cheap — high-volume, low-latency tasks |
Claude is known for being thorough, honest about uncertainty, and genuinely good at following nuanced instructions. The vibes are different from GPT — less eager-to-please, more “here’s what I actually think.”
Products
- Claude.ai — Chat interface (free + Pro at $20/mo)
- Claude API — Developer access (Messages API, tool use, vision)
- Claude for Enterprise — Admin controls, SSO, usage analytics
- Claude Code — Terminal-based AI coding agent. Reads your codebase, plans, writes, tests.
- Model Context Protocol (MCP) — Open standard for connecting AI to tools. The “USB port” for AI capabilities.
MCP is quietly one of the most important things they’ve done. It means any tool built to the MCP standard works with any MCP-compatible AI. No more building custom integrations for each model.
The Research
This is what sets Anthropic apart — they publish safety research that actually advances the field:
- Constitutional AI — Align models with principles, not just human ratings
- Mechanistic Interpretability — Literally mapping what individual neurons do inside models
- Sleeper Agents — Demonstrated that hidden behaviours can survive safety training
- Scaling Monosemanticity — Found interpretable features inside Claude
- Many-Shot Jailbreaking — Discovered and disclosed an attack vector
They publish their vulnerabilities. That’s unusual and important.
Key People
- Dario Amodei — CEO. Former VP of Research at OpenAI. Quiet, technical, mission-driven.
- Daniela Amodei — President. Operations and strategy.
- Chris Olah — Interpretability pioneer. Trying to look inside the black box.
- Jan Leike — Head of Alignment. Left OpenAI’s dissolved Superalignment team in 2024.
The Backstory
Anthropic exists because Dario Amodei and others at OpenAI felt safety wasn’t being prioritised enough as capabilities scaled. They left in 2021 with a core thesis: you need to be at the frontier to study frontier safety. You can’t do alignment research on models from three years ago.
$7.3B+ in funding (Amazon $4B, Google $2B+) means they can compete at the frontier while maintaining the safety focus. Whether the commercial pressures eventually compromise that mission is an open question — but so far, it’s held.
Related
- OpenAI — Where Anthropic’s founders came from
- Constitutional AI — Their alignment approach
- AI Alignment — The broader challenge they’re addressing
- AI Agents — Claude Code as a frontier agent product
- Ilya Sutskever — Another OpenAI departure motivated by safety concerns