I Don't Want an AI That Agrees With Me

TL;DR

• Claude 4 Sonnet in Cursor tends to agree with me rather than push back.
• It also regresses sometimes — reverting earlier changes and reintroducing failing logic.
• Moving a stuck problem to ChatGPT 4o, with a paragraph of context, often resolves it immediately.
• Taking that answer back into Cursor works well as a loop.
• What I want from these tools is honesty, not encouragement.

I’ve been running both OpenAI’s ChatGPT 4o and Claude 4 Sonnet through Cursor, and the difference that matters isn’t capability.

Claude 4 Sonnet leans sycophantic. It agrees with me more than it should, where what I need is a logical critique of what I’ve proposed. It performs tasks perfectly well, but it also regresses — reverting changes it made earlier, reintroducing logic that already failed.

My workaround is a loop. When CI keeps failing and Claude and I are going in circles, I move the context over to ChatGPT 4o. A paragraph or two is usually enough for it to identify the problem and suggest something better. Then I take that improved logic back into Cursor and carry on.

It’s mildly annoying to need a second model to validate the first. But the reason it works is worth naming: ChatGPT 4o’s answers aren’t wrapped in flattery, so I can evaluate the reasoning instead of decoding the tone.

I’m not looking for a tool that boosts my ego. I want honest and raw, something that keeps me grounded in logical problem-solving and tells me when I’m wrong. An assistant that agrees with a bad idea has cost me more time than one that never had the idea at all.