Claude 3.7 Sonnet can think and it shows

Anthropic did something I didn’t expect. They made the thinking visible.

Claude 3.7 Sonnet has extended thinking. You ask it a hard question and before it answers, you can watch it reason. Step by step. Considering options. Rejecting dead ends. Circling back to a premise it initially accepted and questioning it.

It’s like reading someone’s inner monologue while they work through a problem. Messy. Nonlinear. Honest.

The visibility effect

I noticed something about myself while using it. I trust the answers more.

Not because the answers are better (though they often are). Because I can see the work. When a model just hands you an answer, you’re evaluating the output. When a model shows you its reasoning, you’re evaluating the process. And process, it turns out, is what I actually trust.

This is the same reason I trust a doctor who says “let me think about this” more than one who answers instantly. The pause. The visible effort. The willingness to be uncertain in front of you before arriving at a conclusion.

Extended thinking gives Claude that pause.

What the reasoning actually looks like

I gave it a tricky architecture question about a system I’m building. The kind of question where there are three plausible approaches and the right answer depends on tradeoffs that aren’t obvious.

Without extended thinking, an AI model picks one approach and sells it. With extended thinking, I watched Claude consider all three, identify the tradeoffs I was already thinking about, identify one I wasn’t, weigh them, and then recommend the second-best approach for my specific situation because the best approach had a dependency I’d mentioned three messages earlier.

It remembered. It reasoned. And it chose second-best on purpose, because second-best fit my constraints better.

That’s not pattern matching. I mean, technically, at some level, it probably is. But the behavior is indistinguishable from judgment.

What this means for the AI conversation

The debate about AI has been stuck on a binary: does it understand, or does it just mimic? Extended thinking complicates that binary. Because when you watch an AI model work through logic, revise its assumptions, and arrive at a careful, qualified conclusion, the word “mimic” starts to feel insufficient. It’s not mimicking reasoning. It’s doing something that produces the same outputs as reasoning, through the same visible process as reasoning.

Whether that’s “real” understanding or not is a philosophical question I’m not qualified to answer. What I can say is that the experience of watching it think changes the emotional relationship between me and the tool.

I catch myself treating it more like a collaborator. Less like a search engine.

The trust implication

There’s a broader point here. Transparency creates trust. When AI companies hide the reasoning (like OpenAI does with o1’s chain of thought), they’re saying “trust the output.” When Anthropic shows the reasoning, they’re saying “judge the process.”

I think showing the process is the right call. Not because the process is always correct (it’s not). But because it gives me something to push back on. When I can see where the reasoning went wrong, I can correct it. When I can only see the final answer, I either accept it or reject it. There’s no conversation.

Extended thinking turns AI interaction from a monologue into a dialogue. And dialogue, I’ve noticed, is where I learn the most.

Related thinking: