A very late post: this has been sitting in my drafts for months. Most of it reflects how I was using Codex and Claude around Radical Speed Month.

I wrote How I Claude (Intro and Talk to Claude Code) earlier this year because Claude Code changed how I work.

Those posts were true when I wrote them. I still use Claude for code reviews, making plans, and reviewing plans.

During Radical Speed Month, Automattic had free unlimited Codex tokens for a month. The timing was perfect. RSM already pushed me to build faster, test more, and use agents more aggressively, so I had a real reason to try Codex on bigger chunks of work.

That changed my habits.

First it was code reviews. Then small fixes. Then rebases. Then test failures. Now I reach for Codex first for a lot of repo work, especially when I already know the shape of the change and need help grinding through the implementation.

I used Codex on xhigh/fast a lot during RSM. It worked well for planning, implementation, and cleanup loops. I also had Codex and Claude review each other when I wanted a second pass. One agent would write or patch, the other would critique, then I’d run tests and push the next loop.

Codex also feels more straightforward to me. Less fluff, more “here’s the diff.” Claude has started pushing back in ways that don’t always help me. Recently it told me to sleep on a task even though it was the start of the day. I get the intent, but in the moment it felt lazy. I wanted help moving the work forward, not a life coach.

The CI loop mattered. I could keep asking the agent to fix the next failing test until the branch went green. That made TDD easier too. Code felt cheaper to write, so I could ask for the test first, make the failure visible, then patch toward green instead of trying to make the first version perfect.

Claude still feels better to me for some things. I like it for making plans, talking through tradeoffs, and reviewing both plans and code. I also use it when I want a more careful writing tone.

But for the way I work right now, Codex is winning more of the day-to-day coding tasks.

I notice it most on boring work: small refactors, test fixes, rebases, PR cleanup, source checks, and validation. Codex can stay in that loop without needing much ceremony. I can hand it a repo task, watch the diff, run tests, and decide whether to keep going.

I still don’t treat any of this as autopilot. I read the diff. I run the checks. I ask another model to review when the change is risky. The workflow works because the agents are cheap enough to question each other and fast enough that I don’t skip the extra pass.

My model habit changed. I used to think about which model was best. Now I think about which model fits the next 20 minutes of work.

Lately, that answer is Codex more often than Claude.