The question I get asked most often, in various forms: "Should I be using Claude or GPT-4o?"
Wrong question. The right one is: what are you doing right now, in this workflow, at this moment?
I've run both in production for over a year — Claude as the primary driver in Cursor for all engineering work, GPT-4o via API for automation pipelines and anything where latency matters. The setup matters more than the benchmark scores. Here's what I've actually learned.
#For building software, Claude wins. It's not close.
If you're technical and you're using GPT-4o as your coding model inside Cursor, you're paying a penalty you don't have to pay.
Claude 3.5 Sonnet holds more context, generates complete artifacts (not lazy // ... rest of code unchanged placeholders), and pushes back when your approach is flawed. That last part matters more than people admit. Ask GPT-4o to refactor a component tangled with three other files and it'll confidently generate something that breaks your imports. Claude tracks the dependencies. It asks the clarifying question you needed someone to ask.
The architectural understanding is the real difference. And it's an architectural difference — not something a GPT-4o model update will fix.
#For automation workflows, GPT-4o has the edge — but narrower than the hype suggests.
When you're building Make.com integrations or anything requiring strict JSON output, GPT-4o's function calling is more reliable and faster at the API level. At 500 workflow runs per day, that response time gap adds up.
The vision capabilities are also better for document parsing — messy PDFs, mixed-format screenshots, scanned invoices. Claude handles these fine. GPT-4o just has the edge.
But this is a narrow win. If your automation needs complex reasoning (not just structured output), Claude holds up better. Know which problem you're actually solving.
#For writing, Claude is in a different category entirely.
This one is almost unfair to compare.
GPT-4o has a particular prose fingerprint: slightly formal, prone to the same transitional phrases, oddly keen on semicolons. If you've read enough AI-generated content, you start recognizing it on sight. Claude follows tone constraints more precisely. Give it your brand voice, a few examples, and hard rules ("no adverbs, no passive voice, 8th-grade reading level") — it honors them. GPT-4o approximates them.
There's an irony here. If you're using GPT-4o to write content and wondering why it keeps scoring high on AI detectors, the model's stylistic fingerprint is part of the problem.
#The operating setup
Don't pick one. Run both.
Cursor + Claude 3.5 Sonnet for everything code and content. OpenAI API + GPT-4o for background automation and structured data extraction. The model choice per task takes five seconds — it's not worth burning hours on debates that resolve themselves in practice.
The benchmark arguments on X are mostly people who've used each model for 30 minutes debating people who've used each model for 30 minutes. The answer comes from running them on your actual work.