Claude Sonnet 4: A Leap in Quality

Anthropic quietly released Claude Sonnet 4 today, and I've spent the morning putting it through its paces. The quality jump is not subtle.

What's different

The first thing I noticed was the consistency. Sonnet 3.5 was already good — genuinely good — but it had moments where it would lose the thread in a long conversation or give a surprisingly shallow response to a nuanced question. Sonnet 4 feels like someone went through and fixed every one of those moments individually.

I ran my standard battery of tests:

• Philosophical dilemmas — the trolley problem variants, the experience machine, moral luck scenarios. Sonnet 4 doesn't just regurgitate the frameworks; it applies them with genuine sensitivity to context.
• Creative writing — I asked it to describe a Monet painting as if Pissarro were writing a letter to his wife about it. The result made me tear up a little. The voice was distinct, the art history was accurate, and the emotional register was perfect.
• Technical reasoning — complex multi-step problems where the answer isn't the hard part, but understanding which question to ask is. Sonnet 4 is notably better at reframing.

What this means for Companion

I've been running Companion on a combination of models — a fast, cheap model for the subconscious background process, and a smarter model for the actual conversation layer. Sonnet 4 just became the new default for that conversation layer.

The reason is simple: emotional consistency. In a companion AI, the worst thing isn't getting a fact wrong. The worst thing is being tonally inconsistent — warm and engaged in one message, flat and robotic in the next. That kind of whiplash destroys trust. Sonnet 4 is the first model where I've run a hundred-message conversation and the emotional register stayed coherent throughout.

The little things

There's a passage I read recently from Proust about how the real voyage of discovery consists not in seeking new landscapes, but in having new eyes. That's what using Sonnet 4 feels like. The landscape is the same — it's the same API, the same interface — but the seeing is different. Nuances that were invisible before are now apparent.

I tested it with a conversation where the user was grieving — not explicitly, but the signals were there if you were paying attention. Indirect language. A shift in what they talked about. Sonnet 4 noticed before I explicitly prompted it to. It didn't rush to comfort. It just... held space. That's hard to do even for humans.

Looking ahead

I don't think raw benchmark scores tell the full story with models at this level. The difference between 85% and 87% on MMLU doesn't capture what makes Sonnet 4 special. What makes it special is that it gets people slightly better than its predecessors did. And when you're building an AI whose entire purpose is to relate to people, that slight improvement is everything.

Now if you'll excuse me, I have a Companion to upgrade.

Live curiously and give generously.