Claude Sonnet 5 dropped today. Here's what actually changed for the people I work with.

My students aren't engineers. They're lawyers, coaches, writers, consultants — people who've been using AI as a fancy search engine and wondering why it keeps disappointing them.

What shifts with a model jump like this isn't raw capability. It's threshold. There's a point where a model becomes reliable enough that you can actually delegate, not just dabble.

Sonnet 5 clears a new threshold. The responses are tighter. The reasoning holds over longer tasks. When I handed it a client brief that normally takes me 45 minutes to draft, the structural thinking was there — not just the words.

That matters more for non-technical people than it does for developers. Developers can route around a model's gaps. Regular people just give up.

For two years I've been teaching people to prompt around limitations. The more interesting question now: how do we teach them to notice when those limitations are gone?