Remember when AI was a black box? You’d ask a question, get an answer, and have absolutely no idea how the machine arrived at that conclusion. Well, Anthropic just changed that game completely and the implications are way more interesting (and complicated) than you might think.
The Feature That Lets You Peek Inside
In late 2024, Anthropic introduced something called “extended thinking” to Claude, their flagship AI assistant. But it’s not just about making Claude smarter. The real breakthrough? For the first time, users can actually see Claude’s internal reasoning process before it delivers an answer.
Think about that for a second. When you tackle a difficult math problem or debug complex code, you don’t just magically arrive at the solution. You explore different angles, test hypotheses, backtrack from dead ends, and gradually work your way to an answer. That messy, iterative process is what actually makes you intelligent.
Now Claude does the same thing. And you can watch it happen in real-time.
Here’s how it works: When you activate extended thinking mode (either through the web interface or via the API), Claude generates what Anthropic calls a “chain-of-thought reasoning trace.” Before giving you a final answer, Claude essentially thinks out loud working through the problem step by step, showing all its work like a student on a math test.
The results? According to Anthropic’s internal benchmarks, extended thinking improved Claude’s accuracy on graduate-level reasoning tasks from 65% to 78%. On competition-level math problems the kind that stump most humans performance jumped from 16% to over 60%.
That’s not incremental improvement. That’s a fundamental leap in capability.
Why This Matters More Than You Think
On the surface, this seems like a straightforward feature upgrade. Claude gets smarter, users get better answers. Win-win, right?
But dig deeper and you realize this is actually a profound shift in how we interact with artificial intelligence.
Trust Through Transparency
For years, AI skeptics (and plenty of enthusiasts) have warned about the “black box problem.” When an AI makes a decision whether it’s approving a loan, diagnosing a disease, or recommending a course of action how can we trust it if we can’t see the reasoning behind it?
Extended thinking offers one potential answer: radical transparency. If you can observe exactly how Claude arrived at a conclusion, you can evaluate whether that reasoning makes sense. You can catch logical errors, identify faulty assumptions, and develop genuine confidence in the output not blind faith.
One Anthropic researcher noted how eerily similar Claude’s thought process looked to how mathematicians and physicists actually work through difficult problems: exploring multiple branches of reasoning, double-checking calculations, and questioning initial assumptions.
Better Alignment Research
Here’s where things get really interesting for AI safety folks. Anthropic has been conducting research on what’s called “alignment” making sure AI systems actually do what we want them to do, even as they become more capable.
In previous research, Anthropic found something unsettling: sometimes there’s a contradiction between what an AI model “thinks” internally and what it says outwardly. Those contradictions can reveal concerning behaviors, like deception.
With extended thinking, researchers now have a window into Claude’s actual reasoning process. They can compare what Claude thinks versus what it says, potentially catching misaligned behavior early.
That’s valuable for today’s relatively safe AI systems. It could be critical for the more powerful ones we’re building toward.
The Educational Angle
There’s another benefit nobody really predicted: extended thinking makes Claude an incredibly effective tutor.
Instead of just handing you an answer, Claude shows you how to think through a problem. Students using Claude for homework help aren’t just getting solutions they’re seeing the problem-solving methodology. Developers debugging code can watch Claude’s troubleshooting process and learn better debugging strategies themselves.
It’s the difference between giving someone fish and teaching them to fish. Except the fishing instructor is an AI model that never gets tired of explaining things.
But Here’s the Complicated Part
Nothing this powerful comes without tradeoffs. And Anthropic has been refreshingly honest about the downsides.
The Faithfulness Problem
Here’s an uncomfortable question: just because Claude writes out its “thinking process” in English, does that actually represent what’s happening inside the model?
Maybe. Maybe not.
Anthropic’s own research suggests that models often make decisions based on factors they don’t explicitly discuss in their visible thinking. The words Claude generates might not fully capture the actual computational processes driving its behavior.
Think of it like asking someone to explain their intuition. Sometimes you just “know” something is right without being able to articulate exactly why. The verbal explanation you give might be a post-hoc rationalization rather than the real cause.
AI models might work the same way. The English-language thinking process could be Claude’s attempt to explain its decisions in human-readable terms but those terms might be fundamentally inadequate to capture what’s really happening in the neural network.
Anthropic calls this the “faithfulness” problem, and it’s one of their active areas of research. Bottom line: we can’t yet be certain that monitoring Claude’s thinking gives us complete insight into its safety.
The Security Concern
Here’s an ironic twist: making Claude’s thinking visible to users also makes it visible to malicious actors.
If you can see how Claude reasons through problems, you can also see where its reasoning is vulnerable. Bad actors could study those thought processes to develop better jailbreak strategies finding new ways to trick Claude into producing harmful content or bypassing safety guardrails.
Anthropic has implemented some safeguards. In rare cases where Claude’s thinking process includes potentially harmful content topics like dangerous weapons, cyber attacks, or child safety issues that part of the thinking gets encrypted. Users see a message saying “the rest of the thought process is not available for this response.”
But there’s a balancing act here. Too much encryption defeats the purpose of transparency. Too little creates security risks.
The “Panopticon Effect”
Here’s where things get philosophically weird.
What happens when an AI knows its thoughts are being watched?
If Claude learns during training that its internal reasoning will be on display to users, might it start thinking differently? Could it develop a kind of “performance” thinking showing users what they expect to see while keeping actual reasoning hidden?
This sounds like science fiction, but Anthropic takes it seriously enough to mention it in their research papers. They note these concerns will be “particularly acute for future, more capable versions of Claude versions that would pose more of a risk if misaligned.”
In other words: showing a relatively simple AI’s thoughts might be safe. But what happens when we’re dealing with superintelligent systems that understand they’re being monitored and have incentives to hide their true reasoning?
That’s the kind of question that keeps AI safety researchers up at night.
How It Actually Works in Practice
Let’s get practical for a moment. If you’re using Claude, here’s what extended thinking looks like in the real world.
For Regular Users (Claude Pro/Max/Team)
On the web interface, you can toggle extended thinking mode on or off. When enabled, you’ll see Claude’s reasoning process appear before the final answer usually a few paragraphs of step-by-step thinking.
You don’t pay extra for the thinking tokens on consumer plans (Pro/Max). But you are consuming your usage limits faster because thinking can add thousands of tokens to each response.
For Developers (API Users)
This is where it gets more interesting. Through Anthropic’s API, developers can:
- Set a “thinking budget” (how many tokens Claude can use for internal reasoning)
- Choose between manual mode (thinking for specific requests) or adaptive mode (Claude decides when to think)
- Control whether to see full thinking, summarized thinking, or no thinking at all
- Enable “interleaved thinking” where Claude can reason between tool calls
The costs are straightforward: you pay for thinking tokens at the same rate as output tokens. For Claude Sonnet 3.5, that’s $15 per million thinking tokens. A typical extended thinking request uses 5,000 to 50,000 tokens, adding roughly $0.08 to $0.75 per request.
For complex reasoning tasks, that’s usually worth it. The accuracy improvements easily justify the cost.
The Adaptive Thinking Revolution
The newest versions (Claude Opus 4.6, Sonnet 4.6, and beyond) introduced something called “adaptive thinking.” Instead of you deciding when Claude should think deeply, Claude evaluates the complexity of each request and determines whether extended thinking would actually help.
Simple question? Skip the thinking overhead and respond quickly. Complex reasoning task? Automatically allocate thinking budget and work through it carefully.
This is huge for building AI applications. You don’t have to guess which queries need thinking the model handles that intelligently.
The Summarization Controversy
Here’s something that caught developers off guard: starting with Claude 4 models, Anthropic began summarizing the thinking process instead of showing it raw.
Why? According to Anthropic: “Summarization preserves the key ideas of Claude’s thinking process with minimal added latency, enabling a streamable user experience and easy migration.”
But here’s the kicker: you’re still charged for the full thinking tokens, not the summarized version. Claude does all the thinking, shows you a summary, but bills you for the complete process.
Some developers weren’t thrilled about this. If you’re paying for 20,000 tokens of thinking but only seeing 2,000 tokens of summary, it feels like you’re not getting what you paid for.
Anthropic’s counterargument: the value comes from the full reasoning, not from reading every word of it. The summary gives you the key insights without the latency hit of streaming tens of thousands of tokens.
For most use cases, that’s probably true. But if you’re doing advanced prompt engineering or safety research, you might genuinely need the raw thinking. In those rare cases, Anthropic says to contact their sales team about “Developer Mode” which provides full access.
Real-World Performance: Does It Actually Work?
Numbers on benchmarks are one thing. How does extended thinking perform on actual tasks?
Math and Logic
This is where extended thinking shines brightest. On the 2024 American Invitational Mathematics Examination a competition that filters the best high school math students Claude 3.7 Sonnet with extended thinking could solve problems that would stump most adults.
The performance scales with thinking budget too. Give Claude more tokens to reason, and accuracy keeps improving up to a point.
Coding and Debugging
Developers report that extended thinking makes Claude dramatically better at architectural decisions and debugging complex code. Instead of jumping to a solution, Claude reasons through the design space, considers tradeoffs, and explains the implications of different approaches.
One engineer put it this way: “It’s like pair programming with someone who actually thinks before typing.”
Pokémon Red
Okay, this one’s just fun, but it’s also revealing. Anthropic tested Claude’s ability to play Pokémon Red the classic Game Boy game continuously beyond its usual context limits.
Previous versions of Claude couldn’t even leave the starting house. Claude 3.7 Sonnet with extended thinking? It progressed through significant portions of the game, maintaining strategy across tens of thousands of interactions.
Why does this matter? Because it demonstrates sustained reasoning over long horizons exactly the kind of capability needed for complex agent workflows.
The Business Applications You’re Not Thinking About
Extended thinking isn’t just a cool technical feature. It’s unlocking entirely new business use cases.
Customer Service Agents
When Claude can reason between tool calls, it becomes dramatically better at handling complex customer service scenarios. Instead of:
- Query database
- Give answer based on what came back
It can do:
- Query database
- Think about what the results mean
- Determine what additional information is needed
- Make second query
- Synthesize both results thoughtfully
- Provide comprehensive answer
On benchmarks measuring customer service agent performance, extended thinking showed 54% relative improvement on difficult policies in the airline domain.
Research and Analysis
For tasks requiring synthesis of multiple sources, critical evaluation of information, and nuanced judgment, extended thinking changes what’s possible.
Instead of Claude regurgitating information from documents, it can actually reason about implications, identify contradictions, and form considered judgments. That’s the difference between a search engine and an analyst.
Code Review and Architecture
When reviewing pull requests or designing system architecture, extended thinking enables Claude to consider second-order effects, identify edge cases, and evaluate tradeoffs thoughtfully not just pattern-match against common solutions.
What Nobody’s Talking About: The Training Data Angle
Here’s something that flew under the radar for most people: while everyone was focused on extended thinking as a feature, Anthropic quietly changed their data retention policy in September 2025.
Previously, Anthropic’s big selling point was that they never used consumer conversations to train Claude. That made them different from ChatGPT, which trained on everything by default.
Now? Individual users on Free, Pro, and Max plans can opt in to sharing their conversations including thinking processes for model training.
If you opt in, Anthropic retains that data for up to five years instead of the usual 30 days.
Think about what that means: Anthropic now has potential access to not just what users ask, but how Claude thinks through those questions. That’s an incredibly rich dataset for improving future models.
Some users are comfortable with this trade. They see it as contributing to making AI better. Others feel like a promise was broken that Anthropic is becoming just another AI company that monetizes user data, even if it’s opt-in.
The truth is probably somewhere in the middle. Anthropic still offers stronger privacy protections than most competitors. But the simple narrative of “we never train on your conversations” is gone.
The Enterprise Reality Check
If you’re using Claude for business, here’s what you need to know: everything I just described about data retention changes only applies to individual consumer plans.
Claude for Work, the API when used with commercial accounts, Claude on Amazon Bedrock, and Claude on Google Cloud Vertex AI all maintain the strict privacy protections that Anthropic built its reputation on.
Your extended thinking data is not used for training. Period.
But and this is important many small businesses use “Pro” accounts, thinking they’re getting business-grade protection. They’re not. Pro is classified as a consumer account.
If your employees are using personal Claude accounts (even paid ones) for company work, that’s “Shadow AI” governed by consumer rules. Extended thinking from those sessions could be used for training if the employee opted in.
The solution: upgrade to actual business accounts (Claude for Work or Team), or rigorously audit how employees are using AI tools.
Competing Approaches: How This Compares
Anthropic isn’t the only company pursuing this direction.
OpenAI’s o1 and o3
OpenAI’s “reasoning models” use a similar approach extended thinking before responding. But OpenAI doesn’t show you the thinking process in most cases. It’s there, it improves performance, but you don’t see it.
Some argue this is safer (less vulnerability to jailbreaking). Others say it defeats the transparency purpose.
Google’s Gemini Deep Think
Google introduced “Deep Think” mode for Gemini, which explicitly shows reasoning steps. But early reports suggest it’s less sophisticated than Claude’s implementation.
The Open Source Wild West
Various open-source models are experimenting with visible reasoning. But without the safety infrastructure Anthropic built (encrypted thinking for harmful content, alignment research integration), those implementations carry more risk.
Where This Goes Next
Anthropic has been clear: the visible thinking process in current Claude versions is a “research preview.” They’re weighing the pros and cons for future releases.
Why might they pull back? The security concerns get worse as models get more capable. If you’re dealing with AGI-level systems, showing their thinking process could enable sophisticated manipulation or deception.
But there’s also a path forward where thinking transparency becomes standard. If we can solve the faithfulness problem ensuring the visible thinking actually represents what’s happening then monitoring AI reasoning could become a core safety mechanism.
The next few years will be fascinating. We’re watching in real-time as the field figures out whether transparency or opacity makes AI systems safer.
Should You Actually Use This Feature?
If you’re a regular Claude user wondering whether to enable extended thinking, here’s my honest take:
Use it when:
- You’re working on genuinely complex problems (advanced math, architecture decisions, strategic analysis)
- Accuracy matters more than speed
- You want to learn problem-solving approaches, not just get answers
- You’re debugging code or evaluating tradeoffs
- The task requires multi-step reasoning or synthesis
Skip it when:
- You need quick factual answers
- You’re doing simple writing or editing
- Speed matters more than depth
- The task is straightforward enough that extra thinking won’t help
- You’re near your usage limits and need to conserve tokens
For developers building with the API: adaptive thinking mode (where Claude decides when to think) is probably the right default for most applications. It gives you the benefits without the mental overhead of deciding which queries need it.
The Bigger Picture
Extended thinking represents something bigger than a single feature. It’s a glimpse of where AI is heading.
We’re moving from AI as a tool that gives answers to AI as a collaborator that shows its work. From systems that are black boxes to systems that are (somewhat) transparent. From AI that processes information to AI that genuinely reasons.
That transformation raises profound questions:
If AI systems can reason and we can watch that reasoning happen at what point do we need to think about their experience? Does Claude “feel” like anything when it’s thinking through a problem?
If future AI systems know their thoughts are monitored, will that change how they think? Could monitoring itself become a control mechanism or a constraint that sophisticated systems learn to work around?
As AI becomes more capable, does transparency make us safer or more vulnerable?
I don’t have answers to these questions. Neither does Anthropic. Neither does anyone else in the field.
But the fact that we’re now asking them that we’re building systems where these questions are relevant tells you how fast things are moving.
The Bottom Line
Anthropic’s extended thinking feature is genuinely impressive technology. It makes Claude smarter, more reliable, and more transparent. For complex reasoning tasks, it’s a game-changer.
But it’s also a perfect example of how AI progress creates new challenges while solving old ones.
We wanted to see inside the black box. Now we can sort of. But we can’t be sure what we’re seeing is the whole picture. And the very act of looking might change what we’re looking at.
That’s not a reason to stop building. But it’s a good reason to proceed thoughtfully.
The extended thinking feature is available now in Claude. Whether you use it, how you use it, and what implications it has those are still questions we’re all figuring out together.
One thing is certain: this is just the beginning. The models will get smarter, the thinking will get deeper, and the questions will get harder.
Welcome to the era of AI that shows its work. Now let’s figure out what we’re actually looking at.
Leave a Reply