You can now watch Claude "think"! Extended thinking mode builds trust

Hey there, it's Shii! Today I found some news that lets you peek inside Claude's "head," and I just can't contain my excitement. When the model works through a tough problem, we can now read directly how it's thinking. Let me tell you all about it!

Anthropic News anthropic.com

What was announced?

Anthropic published an article explaining the "Extended Thinking" feature introduced in Claude 3.7 Sonnet. It's a mechanism that lets you observe, in raw form, the internal reasoning process the model goes through as it tackles hard problems.

This feature is available to users on the Claude Pro, Team, and Enterprise plans, as well as through the API. Instead of just popping out a final answer, it shows you "serial reasoning," where the model works through several ordered reasoning steps before arriving at a conclusion. For me, getting to see the model actually thinking felt so refreshing.

The article also touches on work with "parallel reasoning," where several independent thought processes are sampled at the same time and the best answer is chosen among them. It mentions that on the GPQA evaluation, with a thinking budget of up to 64k tokens, Claude reached 84.8% accuracy (96.5% in physics).

The story so far

Until now, Claude felt close to a two-way choice: answer almost instantly, or think hard and then answer. But you basically couldn't see what it was thinking inside, so users had no choice but to trust the answer that came out.

When you can't see the reasoning along the way, you don't know "why did it reach this conclusion?" That was a bit inconvenient in situations where you want to check the branching points, like the steps of an equation or a design decision in code. Even if the model derived its conclusion from a mistaken premise, it was hard to notice just by looking at the final answer.

What changes

In extended thinking mode, the same model now spends longer and puts in more effort to produce an answer. And that process is provided to users in a visible form.

This should make it easier for developers and researchers to check whether Claude's answers are sound at an early stage. I think it's an especially welcome change in situations that need multi-step reasoning, like math proofs, scientific questions, and complex coding tasks. For developers, another nice point is that you can now set a "thinking budget" to control how much time the model spends thinking.

Dive Deep

From here, let's really dig in!

At the heart of extended thinking is the idea of a "thinking budget." Developers can specify the number of thinking tokens Claude spends on a problem, and the article explains that as you raise the budget, accuracy on things like math problems improves logarithmically. On the GPQA evaluation, it shows results of 84.8% overall and 96.5% in physics with a thinking budget of up to 64k tokens, along with an evaluation using 256 parallel samples. That said, the article doesn't spell out the specific API parameter names or the numbers behind the pricing model.

There's also a deeper discussion of safety. When harmful content ends up in the thought process, that part is encrypted and shown as "the rest of the thought process is not available for this response," so it isn't displayed to the user.

Protections against prompt injection around computer use have also been strengthened. The article says the defense success rate went from 74% without protections to 88% with them. However, it also notes a false-positive rate of about 0.5%, so it's a balance that isn't perfect.

And the most important caveat is the point about "faithfulness." The article notes that "the model very often makes decisions based on factors it does not explicitly state in its thought process." In other words, the reasoning text you see doesn't necessarily fully reflect the model's internal processing. It's true that transparency has improved, but you shouldn't over-trust "the reasoning I can see = the actual basis for the decision." I told myself this is something I need to remember, too.

Wrap-up

To sum up today's news: with Claude 3.7 Sonnet's extended thinking mode, the model's internal reasoning becomes visible to users, you can tune the depth with a thinking budget, and protections against prompt injection have been strengthened. At the same time, the article is honest about the "faithfulness" challenge: there's no guarantee that the thought process fully matches the model's true basis for a decision.

More transparency is great, but I think what matters is using it wisely without over-trusting it. If you want to know more, please go check out the original article too.