A while back I noticed that a prompt I had used happily for months suddenly gave me a different kind of answer. Not worse, just different. Shorter where I was used to detail, and more precise about what I had literally asked for. It turned out not to be a fluke. Under the hood the model had changed: Claude Sonnet 5 had replaced Sonnet 4.6 as the default.
If you use Claude for work, at home or for your business, it pays to know what is different. Not because your old prompts are broken, they still work, but because a few small adjustments get you noticeably better answers. This first part of the series lays out the three biggest changes. The parts after it go deeper, from beginner to power user.
What Sonnet 5 is
Sonnet 5 is Anthropic’s mid-tier model. Faster and cheaper than the heaviest model, Opus 4.8, but by now good enough that it comes close to that heavy model on hard tasks. For everyday work it is the obvious choice from here on. It is in the free plan, in Pro, and it is the default in Claude Code.
For most people that is all you need to know. What is more interesting is what it means for how you steer it.
Change 1: it now thinks for itself
With the previous version you sometimes had to explicitly ask Claude to think things through before it answered. That was called extended thinking, and you switched it on yourself.
In Sonnet 5 that is gone. The model now adds a thinking step on its own when a question calls for it, and skips it for something simple. Adaptive thinking, it is called. You do not have to do anything for it. In fact, if you still try to pass the old manual thinking instruction, you get an error. It is simply no longer there.
In practice this means you no longer have to worry about whether the model is thinking deeply enough. It handles that itself, based on the question you ask.
Change 2: it follows you more literally
This is the change you will probably feel fastest. Sonnet 5 follows your instructions more precisely than its predecessor. It fills in fewer gaps on its own and guesses less about what else you might have meant.
That sounds like a detail, but it shifts how you write a prompt. You used to be able to be a bit vague and hope the model picked up your intent. Now you get exactly what you wrote. Ask for “a short summary” and you get something short, even if you actually had three paragraphs in mind. The good news: this makes it easier. When you get what you ask for, you know that a disappointing answer comes down to your prompt, not to a guess by the model. The controls are back in your hands.
Change 3: you decide how hard it works
Sonnet 5 has a setting called effort, meaning how much work the model puts into a task. There are five levels, from low to max, with high as the default.
The idea is simple. A quick, simple job you set to a low level, so it goes faster and costs less. A complicated problem you set higher, so it takes more time to reason. The highest levels you save for the exceptions where you really cannot afford a mistake. For anyone who wants to compare: Sonnet 5 at the middle level performs roughly like the previous version at its highest. So you get more, for less.
How to use those levels in practice gets full treatment later in the series.
One caveat
Following you more literally has a flip side. A model that does exactly what you say also carries your mistake through exactly when your instruction is off. The responsibility for what comes out stays with you, not the model. And bear in mind that all of this makes you dependent on a single provider: the more useful it gets, the more of your work is tied to it. That is no reason not to use it, but a reason to do so deliberately.
What you read next
Those are the three big things: it thinks for itself, it follows you more literally, and you decide how hard it works. The rest of this series turns that into something you can use.
In part 2 I start from zero, for anyone who has never really prompted. The anatomy of a good instruction, with examples of wrong and right, and five prompts that always work. In part 3 I go a layer deeper: effort levels in practice, steering the length, and updating your old prompts. In part 4 I close with the approach of the people who use it most heavily.
You do not have to read all of it. Pick the part that fits you. But if you are just starting out, the next piece is where to begin.
Read on in part 2: how to give Claude Sonnet 5 good instructions.