A principal-level engineer with 14 years of experience spent roughly 100 hours co-developing with Claude Code (Opus 4.6) and 20 hours with Codex (GPT-5.4) on the same production codebase: an 80,000-line Python and TypeScript data analysis application with around 2,800 tests, a PostgreSQL backend, WebSocket data streaming, and server-sent events to a web UI. The kind of project where architectural decisions compound and shortcuts show up later as debt.
How the Testing Was Done
Both tools were used with a structured agentic workflow, not open-ended prompting. Each session started in plan mode with a scoped prompt, followed by a multi-subagent review covering architecture, coding standards, UI design, and performance, each referencing explicit documentation built from earlier research sessions. Code was committed per phase, reviewed by specialist subagents, and manually steered based on feedback.
Claude Code: Fast, Productive, Needs a Strong Driver
Speed: Fast and interactive, the strongest advantage Claude has over Codex.
Code quality: Good output overall, but Claude consistently extends existing files rather than creating new ones, producing god classes over time. It adds helper functions instead of revisiting the underlying architecture. Left unchecked, technical debt accumulates in proportion to how fast it moves.
Instruction following: Ignores CLAUDE.md at least once per session. This is a real reliability gap for teams that depend on consistent conventions.
Task completion: Occasionally leaves work half-done, migrating most of a test suite to a new pattern while leaving a handful on the old one. Tests are roughly 95% useful, but the remaining 5% pins broken behavior rather than catching it.
Autonomy: Needs active babysitting. You need to watch the output and intervene. If you are a skilled engineer paying close attention, you will get more done per session with Claude than with any other tool available.
Best for: Rapid prototyping, interactive development sessions, moderate-complexity projects where speed matters.
Codex: Slower, More Deliberate, Closer to Autonomous
Speed: Roughly three to four times slower than Claude on equivalent tasks.
Code quality: Noticeably cleaner. Codex stops mid-task, pulls back, and refactors code without being asked. Where Claude extends god classes, Codex breaks them up. On several occasions it identified improvements the engineer had not considered and implemented them unprompted.
Instruction following: Has never been observed ignoring AGENTS.md. It reportedly will not allow those directives to be overridden mid-session.
Task completion: More thorough. The engineer describes firing off Codex and returning to review finished work, rather than monitoring output line by line to catch problems early.
Autonomy: Operates well unsupervised. Less interactive by nature, which suits fire-and-review workflows better than back-and-forth sessions.
Best for: Production codebases, enterprise software, projects where architectural consistency matters more than raw speed.
The Bottom Line
Claude gets more done per session. Codex work is better. That tradeoff drives everything.
With Claude, you can prototype and build extremely quickly, but expect to guide refactorings every few days and actively manage what it produces. With Codex, progress is slower but the codebase stays cleaner as it scales, and cleanup cycles are less frequent.
For vibe coding or fast prototyping on a low-to-moderate complexity project, Claude is the right call. For building software that needs to hold up over time, Codex has the edge.
One caveat applies to both: neither tool produces good output if the person driving it does not understand software engineering. AI coding assistants amplify the engineer using them, for better or worse.
