Claude Code vs Codex: a Zenn user finds the gap is personality, not performance

The comparison comes from a writer who had long used Claude Code as the main tool and Codex as a reviewer, then upgraded ChatGPT due to dots, an OpenAI resident agent available on Pro 100（月100ドル） and above, so both tools sat at the $100 tier. Codex behaved like an engineer who follows instructions faithfully and checks in at every stage; Claude Code behaved like a flexible engineer who proposes a likely path and moves ahead. The author notes almost no rework from Claude Code's guesses during the month.

The gap narrowed by task. In requirements work and the look of generated apps, Claude Code leaned closer to the author's taste, though the author says deep-thinking settings or a design system can reduce the gap. In technical design and review, Codex flagged architecture, technical debt, and bugs on par with Opus. The pain point came with write operations, where Codex's safety-oriented sandbox blocked commits, pushes, PRs, and updates to progress files outside the repository; both tools can widen writable areas through settings.

It links its view to other writers: tark_ann and arufian align with splitting roles, while toku and junnosukej differ, suggesting where to delegate depends on the person. The author's next-month plan reflects that, returning to Claude Code as the main tool and Codex to a review role. The author credits Codex's frequent usage-limit resets, described in official help as a non-guaranteed benefit, for making full use of the $100 plan during the trial.