So Claude just got a major upgrade. On February 24, 2025, Anthropic released Claude 3.7 Sonnet with a feature they're calling "extended thinking." It's not just marketing fluff – the model genuinely works differently now.
The short version: it can take its time to think through hard problems instead of rushing to an answer. For complex questions, those extra 30 seconds actually matter.
What's Different This Time
Claude 3.7 Sonnet is what they call a "hybrid reasoning model." That's a fancy way of saying it can do two things:
- Quick mode – Works like before. Ask a question, get an instant answer.
- Extended thinking mode – The model actually pauses, thinks through the problem step by step, and then gives you an answer.
This might sound like what OpenAI's o3-mini does, and it's similar. But there's a key difference: Claude 3.7 is the same model doing both jobs. You're not switching to a different AI – you're just giving the same AI more time to think.
Think of it like asking someone a question. Sometimes they know the answer immediately. Sometimes they need to stop and actually work it through. Now Claude can do both.
What It's Good At
Extended thinking improves performance on math, physics, instruction-following, and coding – basically anything where working through the problem step by step helps.
Some practical examples where the extra thinking time pays off:
- Business math – "What's my actual profit margin after I factor in shipping, returns, and credit card fees?" Claude will work through it instead of guessing.
- Debugging code – It can trace through logic errors step by step, catching mistakes it would miss in quick mode.
- Complex writing – Need a proposal that addresses multiple concerns? Extended thinking helps it structure the argument better.
- Analyzing contracts – Spotting problems in long documents where the issues aren't obvious.
The Numbers (If You Care)
Benchmarks are marketing material, but these are worth knowing:
Claude 3.7 Sonnet hit 62.3% on SWE-bench Verified – a test that measures how well AI can solve real software bugs. Claude 3.5 Sonnet scored 49% on the same test. That's not an incremental improvement – it's a significant jump.
The model also reduced unnecessary refusals by 45%. If you've ever had Claude refuse to help with something completely reasonable because it was being overly cautious, that should matter to you.
And here's something interesting: accuracy improves logarithmically the more time you give it to think. More thinking = better answers, but with diminishing returns. There's a sweet spot.
You Can See It Thinking
Here's the part that surprised me: Claude shows you its thought process. Not a summary – the actual messy reasoning it's doing.
OpenAI's reasoning models hide their thinking. Claude shows it. This is actually useful:
- You can catch when it's going down the wrong path
- You can understand why it reached a particular answer
- If it's wrong, you can see where the logic broke down
Anthropic admits the thinking can be "less personal-sounding" than Claude's normal output – it's raw reasoning, not polished prose. But for complex problems, seeing the work matters more than seeing charm.
What It Costs
Same price as before: $3 per million input tokens, $15 per million output tokens. The thinking tokens are included at no extra cost.
For comparison: OpenAI's o1 reasoning model costs $15 per million input tokens and $60 per million output tokens – that's 4-5x more expensive.
If you're using Claude through the web interface, the Pro subscription is still $20/month. Extended thinking is available on Pro, Team, and Enterprise plans – not the free tier.
When NOT to Use Extended Thinking
More thinking isn't always better. Don't use it for:
- Quick questions with obvious answers – "What's the capital of France?" doesn't need 30 seconds of deliberation.
- Casual conversation – Extended thinking mode is more detached. Regular Claude is better for chatting.
- Simple writing tasks – "Write me a short thank-you email" is fine in quick mode.
- When speed matters – If you need an answer in 3 seconds, waiting 30 defeats the purpose.
The model is smart enough to not use the full thinking budget if it doesn't need to. But you'll still feel the delay. Use it when the accuracy actually matters.
For Developers: You Control the Thinking Budget
If you're using the API, this is the interesting part: you can tell Claude exactly how many tokens to spend thinking. Up to 128,000 tokens of thinking – that's over 15 times the output capacity of Claude 3.5 Sonnet.
This lets you build applications where:
- Simple queries get quick answers
- Complex queries get deep thinking
- You control the tradeoff between speed and quality
It's the kind of fine-grained control that matters when you're building something production-grade.
How I'm Using It
My workflow now looks like this:
- Email drafts – Regular Claude. Fast, good tone.
- Complex proposals – Extended thinking. Better structure, catches edge cases.
- Code debugging – Extended thinking. Worth the wait when you're stuck.
- Quick research – Regular Claude or ChatGPT with web access.
- Contract review – Extended thinking. Catches things I'd miss.
- Creative brainstorming – Regular Claude. The detached thinking mode is less fun here.
The key is matching the tool to the task. Extended thinking is a power tool – excellent when you need it, overkill when you don't.
Claude vs ChatGPT vs o3-mini: Where We Are Now
Here's my current take on which AI to use for what:
- Business writing that needs nuance → Claude 3.7 Sonnet (regular mode)
- Complex reasoning or math → Claude 3.7 (extended thinking) or o3-mini
- Web research → ChatGPT (still the only one that browses)
- Image generation → ChatGPT with DALL-E
- Long document analysis → Claude (best context window)
- Coding tasks → Claude 3.7 (extended thinking) – the benchmarks back this up
If you already use Claude, you're getting a genuine upgrade at no extra cost. If you're deciding between AI subscriptions, Claude Pro just got more compelling.
The Bottom Line
Claude 3.7 Sonnet is a real improvement, not just marketing. The extended thinking feature genuinely helps with complex problems. The visible reasoning helps you trust (or question) the answers. And it's cheaper than OpenAI's reasoning models.
My recommendation:
- Try extended thinking on a real problem – Something you know the answer to. See if the extra time gives better results.
- Use it for complex tasks – Math, coding, analysis. That's where it shines.
- Don't use it for everything – Quick questions don't need 30 seconds of deliberation.
- Watch the thinking – Reading Claude's reasoning is actually useful for catching mistakes.
The 30-second wait takes getting used to. But when you're dealing with something that actually matters, seeing the AI work through it properly is worth the time.
What YouGrow Does Differently
We use Claude every day for client work – and honestly, Claude is my favorite AI tool. Extended thinking makes it even better for the complex stuff: analyzing competitors, debugging custom tools, writing proposals that actually address the real concerns.
Not sure which AI tools make sense for your business? Let's talk. I'll give you an honest opinion on what's actually worth your time – no upsell, just straight answers. If you're looking for a managed website service, check out YouGrow too.