Claude vs Codex：本番コードベース比較

an engineer with 14 years of experience with time at a MAG7 company spent 20+ hours this weekend running both tools on a real production codebase. here's what he found (you might be changing your stack after reading this article): I came across this post on reddit (linked at the end) that compared the two, it was interesting because it wasn't from the perspective of a "vibe coding hobbyist" it was from the perspective of a engineer that is working at a major tech firm at principal/staff eng manager equivalent with prev MAG7 experience. THE SHARED WORKFLOW (BOTH TOOLS RAN THE SAME RULES) before getting into the results, the workflow matters. both tools operated under the same system: plan mode first - a thorough, scoped prompt, then an 8-subagent plan-review skill fires. architecture review checks SOLID, DRY, KISS, YAGNI with explicit reference documents (postgres_performance.md, python_threading.md, software_architecture.md). each subagent has tightening prompts. code phase - each plan phase committed separately, code-review skill runs on each commit, output manually reviewed and steered. CLAUDE.md / AGENTS.md both sitting at ~100 lines - TDD rules, git workflow, key devex conventions, docker commands. same driver. same rules. very different behaviour from the tools. WHAT CLAUDE ACTUALLY DOES ON A REAL CODEBASE claude feels like a senior engineer on a time crunch. smart. fast. genuinely capable. but trying to ship the feature and move on, not worrying too much about whether the architecture is clean on the other side. a few observations from someone who has been running it on an 80k LOC codebase for months: the 1MM context window is a noob trap. keeping it under 250k produces far more predictable behaviour. and even then - claude ignores CLAUDE.md at least once a session. conventions explicitly written down. gone. like they weren't there. it almost never creates new files for new functionality. it reaches for existing files and keeps extending them. for an engineer who comes from C/C++ and wants files under ~600 lines with strong separation of concerns, this is a daily friction point. give it an 8-suite test migration task and it will complete most of them and quietly leave the rest on the old pattern. you have to check. the test problem is the one that compounds. left unchecked, claude changes tests to match what it thinks the goal is. 95% of the tests it writes are solid. 5% pin broken behaviour. over time, that ratio matters. the interactive speed is excellent for prototyping. but it needs a skilled, focused driver with eyes on the output at all times. WHAT CODEX ACTUALLY DOES ON A REAL CODEBASE codex feels like a solid 5-6 year engineer. not junior. not principal. but genuinely thoughtful. here's where it gets interesting. on multiple occasions, codex stopped mid-task, recognised it was extending a god class, and refactored the whole thing into clean components without being prompted. claude would have dropped another function at the bottom of the file and called it done. it runs 3-4x slower than claude for the same task. that is not a small gap. but it doesn't need babysitting. the engineer's workflow became: fire it off, come back when it's done, review the work. it demonstrated enough competence that watching the output line by line felt unnecessary. it has never once ignored AGENTS.md. it won't let you override directives mid-session. that level of instruction fidelity changes how you build. a handful of times it produced improvements the engineer hadn't thought of. not hallucinations. genuinely additive decisions. the autonomy is real. the slowness is real. you are paying for one with the other. THE HONEST OVERALL VERDICT codex pro x5 has roughly similar usage caps to claude x20. worth knowing before you budget. in a session with claude, more gets built. prototyping is fast. but every few days there is a refactoring pass to clean up accumulated cruft that the tool introduced. with codex, refactoring passes still happen, this is driven by natural app growth rather than "let me go see what mess needs fixing." for prototyping and fast-moving low-to-moderate complexity projects: claude. for enterprise software that has to live and scale: codex. and this is the part that doesn't get said enough: both tools produce garbage if you don't know software engineering. they are multipliers. bad inputs get multiplied faster, not fixed. WHAT THIS MEANS FOR YOUR STACK the framework any senior engineer should carry into this decision: 1. match the tool to the phase. claude for exploration and prototyping. codex when you're building something that has to last. 2. invest in your .md files. CLAUDE.md and AGENTS.md are the difference between a powerful tool and a liability. 100 lines of well-crafted conventions pays back every single session. 3. build a plan-review layer. 8 specialised subagents checking architecture, coding standards, performance, and UI design before a single line gets written is not overkill. it's the workflow that makes both tools production-grade. 4. guard your tests. ai coding tools will drift your test suite if left unchecked. the rule that works: if tests break after a change, stop and flag it. do not let the tool fix them silently. the best ai coding setup is the one you can actually govern. which of these are you running in your stack right now? claude, codex, or something else? You can now see why @steipete loves Codex so much. Here's the link to the original post: https://www.reddit.com/r/ClaudeCode/s/1QpTzC54vQ