10 Comments
User's avatar
Alex Lastovetskiy's avatar

Useful breakdown. The evaluation methodology angle especially — that's where the gap between theory and production usually shows up.

SOMEONE's avatar

Protection against destructive commands needs to be ensured by the harness and the sandbox. You can nudge the model towards it but ensure it there.

Victualis's avatar

If different systems A and B are reported as working roughly equally well by different people, it could be that they are roughly equivalent in capability but stochastic effects are real. OTOH I am often surprised by some of the difficulties people report. I seem to have found the model that is in tune with how I vibe and I'm astounded daily by the great work it does. This is maybe evidence that there really is a "match" based on user personality and habits.

Pierre Brunelle's avatar

I have the impression some people over index based on very little experience (eg one model is better than another on a single problem). I personally can’t really tell the difference between Sonnet 4.5 and Opus 4.5 (now 4.6) despite using each of them about 30 hours per week. However aggregating the feedback over many reactions (like Zvi does) gives a useful picture.

me-AI's avatar

It's intriguing to see the development of models like Codex and how they enhance coding capabilities. If you're interested in the underlying mechanics of AI's thinking, you might find the exploration of language model interpretability in my recent piece enlightening: https://00meai.substack.com/p/brain-damage-teaches-us-how-language.

Matthias U's avatar

Why the heck do they make using 5.3-Codex so damn difficult? I seem to have no way to access it. "The model `gpt-5.3-codex` does notexist or you do not have access to it.)"

Pawel Jozefiak's avatar

Been using both daily since launch. The competition is pushing both tools forward fast.

Codex's speed advantage is real for iteration. But Claude Code's reasoning depth and frontend design instincts still win for anything beyond quick prototyping. The "also good" framing is accurate—it's not "better," it's "differently good."

Wrote a head-to-head comparison after 2 months of daily use: https://thoughts.jock.pl/p/claude-code-vs-codex-real-comparison-2026

The interesting part isn't which model wins. It's how the ecosystem around each shapes the actual developer experience.

Alex Lastovetskiy's avatar

Good analysis. The decision-making frameworks here apply beyond the specific topic.

JP's avatar

The safety rating escalation is the part that grabbed me. The CLI's sandbox modes actually address some of this. You can lock it down so every file write needs sign-off, or go full auto if you trust the scope. Most guides skip the safety architecture entirely so I wrote it up properly: https://reading.sh/the-definitive-guide-to-codex-cli-from-first-install-to-production-workflows-a9f1e7c887ab