In 2021, Dario and Daniela Amodei left OpenAI. They disagreed with the direction of the company — specifically around safety and the pace of deployment. They took eleven colleagues with them.
They raised $124M in seed funding and named the company Anthropic. The explicit mission: AI safety research. Build slowly. Build carefully. Publish your methods. Don’t rush to capabilities.
The unexpected result: they built the best coding AI in the world.
On the SWE-bench Verified benchmark — a dataset of real, unsolved GitHub issues from major open source repositories — Claude 3.5 Sonnet resolved 49% of issues. GPT-4o resolved 38%. The gap was not marginal.
Cursor, the AI coding editor that became the fastest-growing dev tool of 2024, defaults to Claude Sonnet as its primary model. When millions of developers ask AI to write production code every day, they’re mostly talking to Anthropic’s model.
Why does a “safety-focused” lab build the best coding AI?
The answer is less surprising than it seems. Anthropic’s core research bet was Constitutional AI — training models to reason carefully about instructions, check their own outputs against a set of principles, and refuse requests that violate them.
A model that reasons carefully about instructions writes better code. A model trained to check its own outputs catches its own bugs. The same properties that make Claude safe also make it accurate.
It turned out that “safe AI” and “capable AI” weren’t opposites after all.
Anthropic has raised over $7.3 billion — from Google ($2B), Amazon ($4B), and Spark Capital. The company is now valued at $61 billion, making it the third most valuable private company in the US behind SpaceX and Stripe.
Dario was right about both things. You can build carefully and win on performance. The two goals reinforce each other.
The people who said safety and capability were a tradeoff were wrong. And now they’re catching up.