You might ask why Anthropic, and it’s a question worth asking.

First, the origin story.

🎯 The Founder’s Journey

The founder of Anthropic, Dario Amodei, is somebody I can trust. He was one of the lead Engineers when GPT-2 and GPT-3 were built by OpenAI. He along with his sister and a few more supporters split away from OpenAI because they did not like the direction OpenAI had taken. They felt that OpenAI was drifting away from its original purpose of being a non profit organisation that was interested in building safe AI. His main objection was a difference in vision about how to build safe AI and at what pace. In his own words, “It is incredibly unproductive to try and argue with someone else’s vision”.

🔐 Safety as Foundation

So the foundation of Anthropic is to create safe AI. It is a very serious concern for the founding members, so much so that they put their careers at risk and moved away from an obviously growing and renowned company like OpenAI. Right from the beginning they focussed on how to build a safe AI. They wrote a constitution and ensured that any models they release adhere to this constitution. When others were doing anything they can to be profitable, this team willingly chose a harder path for a principle they believed in.

🏢 Enterprise Focus from Day One

Because they started with this commercial handicap, they are ultra focussed in what they do. Right from day one they only built the model for enterprises, not general consumers. They were trying to build something valuable in itself, not for consumer attention.

💻 Coding Excellence

They also chose one use case of building the best coding model possible. They did not go after creating images or videos or all the other things all the other models are trained to do. For a long time the Claude model seemed very weak in comparison to other models, but it was steadily getting better at code. Sonnet 3.7 was when it became very clear to almost everyone in the developer community that here is a special model. All leading coding assistants (Copilot, Cursor, Windsurf etc) provided Claude Sonnet 3.7 as their most heavily used model.

Since then, the model kept moving away from the pack. Their models are always on the top of SWE-bench and other benchmarks. But more than those league tables, there’s some magic in their way of working with code that other models simply couldn’t match up. Just in the last month they released Sonnet 4.5 which was clearly better than their previous Sonnet 4. Every new version is clearly better than the version it replaces. Just recently they released Haiku 4.5 which is quite powerful, fast and cheap for their entry level model.

Now that I’ve already got a couple of posts out, my writing bone is tickled. All of this post is painstakingly written by me and Claude is relegated to just checking grammar, clarity and fact checking, which it does way better than I can, but obviously it’s sulking at its relegation to this secondary status!!