Quick impressions: A week of using Codex more than Claude

Codex is preferred for users seeking a technical, contained tool that adheres strictly to instructions without over-engineering.

Key Points

Codex acts as a precise companion that follows explicit instructions and stops once the task is complete, whereas Claude is more proactive, guessing user intent and often going beyond the initial request.
In terms of code architecture, Codex produces simpler solutions with fewer abstractions; conversely, Claude creates more complex structures (e.g., Sorbet signatures, type aliases) that typically handle more edge cases.
The user experience differs in ‘personality’: Codex feels like a highly technical entity (compared to Data from Star Trek), while Claude feels more like a collaborative colleague in a pair-programming session.
Codex is perceived as faster for initial changes and produces cleaner Ruby/Rails code with fewer comments, though it may require more manual verification and testing to complete a pull request.
Codex handles MCP authentication more reliably via a CLI approach (codex mcp login), whereas Claude’s automated attempts can sometimes hang.
Codex exhibited weaknesses in complex git operations, such as incorrect rebasing logic when dealing with multiple target branches, requiring explicit manual correction.