The Setup: Why I Actually Did This
Look, I wasn’t trying to be a tech contrarian. I genuinely just got tired of ChatGPT feeling like the default answer to every AI question. You know how it is when something becomes the obvious choice? You stop actually choosing. So when Anthropic dropped Claude 3.7 Sonnet in February, I decided to spend a month using it for everything I normally outsource to Claude’s competitor. Writing prompts, coding help, research, the works. No switching back. Full commitment, even when it felt inefficient.
The timing felt right. Anthropic’s official Claude 3.7 Sonnet announcement highlighted something that caught my attention: extended thinking capability. The model now shows you its reasoning chain before giving you an answer. It thinks out loud. That felt worth investigating for a full month.
I should mention upfront that I’m not a developer, so this isn’t a technical deep dive. This is a regular person who writes, researches, and occasionally needs help debugging why their Python script is throwing inexplicable errors. That’s the lens here.
The Extended Thinking Thing Actually Changed How I Work
Here’s the weirdest part: once I got used to seeing Claude’s reasoning process, I couldn’t go back to just getting an answer. It sounds small, but it’s not. When you watch the model work through a problem, you can catch where it’s making assumptions. You can see if it’s heading down a wrong path before it commits to the final answer.
I used this most with research projects. I’d ask it to help me understand a complicated topic, and instead of just getting a summary, I’d see it wrestling with the source material, identifying contradictions, weighing different interpretations. The extended thinking output is sometimes longer than the actual answer. Most people probably skip it. I started reading it first.
The downside is real though. It’s slower. If you’re used to ChatGPT’s instant gratification, waiting for the model to show its work feels glacial. On my laptop, some extended thinking requests took 30-45 seconds. That’s not long in absolute terms, but it changes the rhythm of how you work. You can’t rapid-fire prompts. You have to actually wait and read.
The Coding Part Was Where I Noticed the Gap
Full transparency: I can’t evaluate coding performance the way an actual developer would. But I can notice when something works and when it doesn’t. Over the month, I threw various coding problems at Claude 3.7 Sonnet. Simple stuff mostly. A JavaScript function that wasn’t parsing correctly. Python scripts that needed optimization. Regex patterns that made my brain hurt.
It felt noticeably more reliable than I expected. I had fewer instances of Claude confidently suggesting code that simply wouldn’t run. When it did make mistakes, they were usually the kind I could spot and fix quickly, not the kind that sent me down a 20-minute rabbit hole.
Anthropic published some internal benchmarks on this. On something called SWE-bench Verified, which tests reasoning about software engineering, Claude 3.7 Sonnet scored 70.3% compared to GPT-4o’s 38.8%. That’s a meaningful gap. The Stanford HAI 2025 AI Index Report noted that reasoning benchmarks across top AI models have narrowed to under 5% difference by mid-2025, but apparently there’s still variance in how that plays out on specific tasks.
The honest part? I can’t tell if Claude’s advantage here was real or if I just had lower expectations coming in. Both are possibilities worth sitting with.
The Context Window Thing Is Actually Useful, Not Just a Parlor Trick
Claude’s context window is 200,000 tokens, roughly equivalent to a 500-page novel. You can dump entire documents into a single prompt and ask the model to synthesize, critique, or reorganize them. I tested this more out of curiosity than necessity, but it turned out to be genuinely helpful.
I had a messy email thread with about 20 different people discussing a project. Instead of manually summarizing it, I pasted the whole thing into Claude and asked for a timeline of decisions and who made them. It worked. It actually worked. That’s the kind of task that feels like it should fail but doesn’t, and once you realize it won’t fail, you start using it differently.
The caveat is that bigger context windows don’t automatically mean better results. You still need to ask good questions. But they do remove a barrier to uploading complex material. With ChatGPT, I find myself pre-editing and summarizing documents before pasting them. With Claude, I’m more likely to just dump the full thing in and see what happens.
What I’m Still Not Sure About
Here’s the thing about spending a month with one tool: you start to miss the other one, but you’re not sure if you’re missing it for the right reasons. By week three, I found myself wondering if ChatGPT felt better because it actually was better or because I was just used to its particular flavor of helpful.
I noticed Claude 3.7 Sonnet sometimes felt overly cautious. More disclaimers. More hedging. More “I should note that this is just one perspective.” That could be a feature or a bug depending on what you’re doing. For personal projects, it felt a bit much. For research where you need accountability built into the answers, it was fine.
The financial context matters too. Anthropic raised 2.75 billion dollars in early 2025, bringing the company’s valuation to 61.5 billion. That’s real capital behind real ambition. But it doesn’t tell you whether Claude or ChatGPT will be better for your specific needs six months from now. The AI landscape is moving too fast for that kind of certainty.
What I’m left with is this: Claude 3.7 Sonnet is genuinely good. It has real advantages, particularly in reasoning transparency and coding reliability. It’s also not a revelation. It’s a solid tool that matters more the more you actually use it instead of just comparing benchmarks.
If you’re considering making the switch, the honest answer is: try it for a week. See if the extended thinking approach clicks for you. See if you prefer the different cautiousness. There’s no objectively correct choice here, just different tools for different ways of thinking.
What’s been your experience if you’ve tried it? I’m genuinely curious whether the extended thinking feature resonates with other people or if it felt like a gimmick to you.
