I've been unusually quiet here lately — head down building a couple of AI apps, which has meant dozens of hours at the keyboard over the past fortnight.
Naturally, I've been using AI to help write these apps. And because I've been hitting usage limits regularly, I've ended up switching between Claude Code and Codex (OpenAI's developer tool) multiple times daily.
Until now, I've been hesitant to recommend one vendor over another. The landscape moves too fast, and what works for me might not work for your stack or your workflow.
But after two weeks of constant switching? I'm ready to put a stake in the ground: Claude Code is materially better for my work.
The telling moment: I found myself topping up my API account so I could keep using Claude Code after hitting my Pro subscription limit, rather than just switching back to Codex. That's not a decision I made consciously — it's what my fingers did when I wanted to keep working.
A few observations…
The Wrapper Matters More Than the Model
Yes, I'm using top-tier models (GPT-5 Codex and Sonnet 4.5), but I'm increasingly convinced the bigger differences come from the products that wrap these models. The UX, the autonomy settings, how they handle context — that's where the real productivity gains live.
Claude Code's Strengths
The killer combination is speed plus autonomy. Claude feels like pairing with another developer — it works fast enough that I can stay engaged in the flow, and autonomous enough that I'm not constantly approving trivial operations. I can actually focus on other things whilst it's working.
Codex feels like babysitting. I can't trust it to work unsupervised, which means I'm stuck watching it. But it's not fast enough to keep me fully engaged, so I end up half-paying attention whilst doing a Sudoku in another tab, waiting for the inevitable moment when it needs me to intervene. It's the worst of both worlds.
For my recent .NET work, Claude was also far superior with CLI tools like dotnet, gh, and az.
Where Codex Still Edges Ahead
Workflow adherence. I've got specific instructions in my AGENTS/CLAUDE.md about branching and committing, and Claude will occasionally forget to create a new branch or try to commit before I've reviewed. Minor annoyance, easily caught, but Codex tends to stick to the script better.
Cost efficiency also favours Codex — the research suggests it's significantly more token-efficient, which could matter at team scale. As a solo developer, I haven't found this particularly relevant. The large-scale Claude Code deployments I've overseen have had overwhelmingly positive ROIs, so the token costs haven't been the limiting factor. But I can see how the economics might shift for teams running at higher volumes.
The Pattern That Interests Me
The general sentiment online suggests Claude dominates frontend and Codex owns backend. My experience doesn't quite match that neat division. Claude absolutely sorted a gnarly SignalR integration into my frontend that Codex was making a mess of. But then Claude also figured out a backend race condition in my event-based architecture that had Codex stumped.
What I'm seeing is less "Claude = FE, Codex = BE" and more that they have different reasoning patterns. When one gets stuck, the other often finds the path through. Overall though, I find I get better results from Claude Code.
The Broader Point
Tool choice matters, but it's not binary. I'm keeping both subscriptions active because they each have their place. The key is having enough experience with the tools to know which one to reach for in different situations.
A Word for Teams
I'm (currently) a solo developer, which makes switching between tools trivial. For teams, it's more complicated. You need to be deliberate about which tools you're using and why. The switching costs at a team level — different workflows, different mental models, different edge cases — are going to be higher than any marginal performance gains.
Don't leap onto whichever tool seems best this week. Pick something, get good at it as a team, and only switch when there's a compelling reason.
That said, for solo developers I'd absolutely encourage trying Claude Code for some meaningful work and comparing it to whatever you're currently using. And for teams not yet settled on a tool, I'd definitely recommend taking a look at Claude Code.
What's your experience been? Are you finding meaningful differences between the various AI coding tools, or does it all feel much of a muchness?