Claude Code, Codex, Cursor: which one are you actually using day to day?

Switched to Claude on a friend's recommendation and stuck with it. What kept me is it actually pushes back on my ideas instead of just agreeing with everything, other chatbots I've tried feel like yes-men in comparison.

I'd been using Codex and switched to Claude Code this week. It is too early to tell, but so far I have more success with Claude Code.

Claude and ChatGPT are both open on my machine right now. They replaced a lot of the manual work—and countless searches through Stack Overflow.

We use both OpenAI and Anthropic because different parts of autonomous testing require different strengths. Some tasks benefit from speed, while others require deeper reasoning.

Our philosophy is governed autonomy: AI agents handle the heavy lifting.

Where I still don't fully trust AI? Customer-facing content without human review. It's an incredible accelerator, but judgment still matters.

Claude Code, open pretty much all day at this point. replaced Cursor for me for anything past a single-file tweak, mostly because I can describe the outcome and walk away instead of babysitting a diff. still keep Cursor around for quick visual CSS stuff where seeing it live matters more than describing it.

Claude Code is what's actually open here too, so the convergence pattern in this thread checks out from where I'm sitting. What's more interesting than which tool won is that teams land on it without coordinating, which usually means it fits how the work already gets talked about between people, not just how it benchmarks solo. The trust answers are consistent as well, migrations, auth, anything irreversible stays manual no matter how good the model gets. That's probably the more durable signal than any benchmark here.

Claude is what stays open here too, but the interesting replacement was not another model so much as the old loop of search, docs, and tiny glue-code detours. The place I still do not trust any of them is the same place users stop trusting phone agents: auth, migrations, billing, identity, anything that changes state with a blast radius. Once a tool is good, the real benchmark stops being output quality and becomes whether the approval boundary moves.

My answer is increasingly: the harness + continuity matter more than the model.

Day to day I keep ChatGPT open for long reasoning/review, use Codex for deliberate coding turns, and bridge back to the Mac so the chat can inspect files/terminal/PTY and read prior Codex sessions. That replaced a surprising amount of copy/pasting between tools.

The trust line is still execution evidence: I want to know what files were inspected, what commands actually ran, the exit codes/tests, and whether the repo changed. A fluent “done” message isn’t evidence. So I switch models less than I switch execution context now.

Claude for personal, codex for work. it seems we need a basic understanding of all of the "top tools" capabilities and limitations. thankfully the UI isn't so different from tool to tool, mainly the output is the biggest differentiator. "can it do this task well?", i have entered the same prompt into the same tool and gotten different responses; i've also put in the same prompt into different tools and got weirdly similar results. we're paying more and more but compromising for 'good enough'