Comment by lhl

As some other people mentioned, using both/multiple is the way to go if it's within your means.

I've been working on a wide range of relatively projects and I find that the latest GPT-5.2+ models seem to be generally better coders than Opus 4.6, however the latter tends to be better at big picture thinking, structuring, and communicating so I tend to iterate through Opus 4.6 max -> GPT-5.2 xhigh -> GPT-5.3-Codex xhigh -> GPT-5.4 xhigh. I've found GPT-5.3-Codex is the most detail oriented, but not necessarily the best coder. One interesting thing is for my high-stakes project, I have one coder lane but use all the models do independent review and they tend to catch different subsets of implementation bugs. I also notice huge behavioral changes based on changing AGENTS.md.

In terms of the apps, while Claude Code was ahead for a long while, I'd say Codex has largely caught up in terms of ergonomics, and in some things, like the way it let's you inline or append steering, I like it better now (or where it's far, far, ahead - the compaction is night and day better in Codex).

(These observations are based on about 10-20B/mo combined cached tokens, human-in-the-loop, so heavy usage and most code I no longer eyeball, but not dark factory/slop cannon levels. I haven't found (or built) a multi-agent control plane I really like yet.)