I love a good head-to-head test. Not the marketing kind where everything wins. The kind where someone actually uses both tools for days and comes back with scars.
That’s exactly what I found in this LinkedIn post. The author spent three days running GPT-5.6 across ChatGPT Work and Codex, side by side with Claude Code. Then he wrote down what actually happened. No hype, no vendor loyalty, just a real workday comparison.
And his verdict surprised me a little.
The one-line summary from the author
“GPT-5.6 is your best employee when you know exactly what you want. But it rarely goes the extra mile for you.”
That’s the whole review compressed into two sentences. It’s an excellent worker. It’s just not a proactive one.
The original poster added a detail that made me laugh out loud. When GPT-5.6 makes a mistake, it tells you “You are right” so you always feel correct. His reaction: (wait, am I?)
I think that’s a sharper point than it looks. A tool that agrees with you constantly isn’t checking your work. It’s flattering you.
The dashboard test
Here’s the story that sold me on his argument.
Before lunch, the expert handed Codex a task: build a dashboard from his LinkedIn data.
- After the first prompt, 80% of the dashboard was already perfect
- The remaining 20% ate the entire rest of his day
- After dinner, he was still fixing details in Codex
Meanwhile, on the same day, his Claude Code had already prepared a full presentation, handled admin work, and fixed issues on a website project.
Same hours. Wildly different output.
Where GPT-5.6 shines
The creator is clear that this isn’t a hit piece. GPT-5.6 is genuinely strong at building things. The 80% it delivered on the first prompt is impressive by any standard.
GPT-5.6 wins when:
- You know exactly what you want
- You know exactly how it should be done
- You can describe the process step by step
- You can name the specific tools to use
Give it a precise spec and it executes fast. That’s a real strength.
Where it stumbles
The gap shows up in the details. According to this industry pro, the model may ignore very basic things: text alignment, aesthetics, the small polish that separates “built” from “finished.”
That last 20% is where the hours go. And it’s the part nobody puts in the demo video.
The tradeoff, plainly:
- GPT-5.6: fast to 80%, slow and fiddly on the final 20%
- Claude: better at picking up the details you didn’t spell out
The author’s recommendation
He lands somewhere refreshingly practical. It’s not “tool X is better.” It’s “match the tool to how much you’ve already figured out.”
- If you know what you want, how it should be done, can describe the process, and can name the tools, use GPT-5.6
- Otherwise, use Claude Opus or Fable within the Claude ecosystem of tools and skills
Why it matters: the real difference isn’t raw model quality. It’s how much thinking you’re offloading. GPT-5.6 wants a finished blueprint. Claude is more comfortable filling gaps you left open.
How you can use this today
You don’t have to pick a side. Pick a mode.
- Spec is locked? Reach for GPT-5.6 and let it sprint
- Still figuring it out? Reach for Claude and let it think alongside you
- Building something visual? Budget real time for that last 20%, whichever tool you choose
- Getting agreed with a lot? Treat “you are right” as a warning sign, not a green light
The bit I keep coming back to is the agreement problem. We’re all a little vulnerable to a tool that never pushes back. Worth remembering next time your AI enthusiastically confirms your worst idea.
Go read the full LinkedIn post for his complete breakdown. The dashboard story alone is worth your time.