Quick impressions: A week of using Codex more than Claude
This post offers quick, personal impressions comparing a week of using OpenAI's Codex more than Anthropic's Claude as a coding assistant. The author highlights differences in code quality, interaction style, and architectural output, sparking a vibrant Hacker News debate on the 'personalities' and practical utility of various LLM-powered developer tools. The discussion frequently diverges into broad critiques of model verbosity and the rapidly evolving landscape of AI coding assistants beyond just the two featured.
The Lowdown
Lucian Ghinda shares his initial, subjective experiences after shifting his primary AI coding assistant from Claude to Codex for a week. He focuses on practical observations related to workflow, code output, and interaction styles between the two prominent models.
- Skill Porting: Claude initially had more custom skills, requiring a porting effort to bring Codex up to parity.
- Familiarity Bias: In urgent debugging scenarios, the author instinctively reverted to Claude due to its familiarity, despite no clear performance advantage.
- Code Comments: Codex generated cleaner code with fewer comments, which the author preferred, contrasting with Claude's tendency for verbose annotations.
- Interaction Style: Codex felt more 'technical' and 'Data from Star Trek'-like, while Claude offered a more conversational, 'colleague in a Tuple session' feel.
- Session Management: The author found himself preferring many focused Codex sessions over single, large Claude sessions.
- Speed vs. Thoroughness: While Codex made changes faster, the overall time to complete a pull request was similar due to its thorough testing and review processes.
- Architectural Simplicity: Codex tended to produce simpler code architectures, whereas Claude often introduced more abstractions, concepts, and type signatures.
- Branching & Rebasing: Codex struggled with complex Git operations like rebasing when not given explicit instructions, leading to large, problematic pull requests.
- Tool Integration (Jira/Atlassian): Codex had a more cumbersome experience with Jira via CLI, while Claude was more adept at integrating within the author's preferred environment.
- Authentication (MCPs): Codex's explicit
codex mcp loginflow was preferred over Claude's sometimes-stuck automatic authentication attempts.
In conclusion, Ghinda posits that Claude attempts to anticipate and 'overdo' tasks by guessing user intent, while Codex acts more as a disciplined companion, executing instructions without over-elaboration.
The Gossip
Model Mastery and Mismatched Methods
Commenters offered wildly contradictory experiences regarding Codex's and Claude's tendencies for complexity and understanding user intent. While the author found Codex simpler, many users reported the opposite, citing Codex's over-engineering and 'relentless' nature, especially with its Sol model. The discussion underscored that performance heavily depends on the specific model, harness, and task, with many users having developed nuanced workflows for which tool best suits which problem. Beyond the two main contenders, a rich ecosystem of other models like Fable, Gemini, Kimi, Grok, and local solutions were frequently brought up as viable and sometimes superior alternatives.
Verbose Verdicts: Claude's Commenting Conundrum
A dominant theme in the comments was frustration with Claude's verbose and often unhelpful code comments. Many users described Claude's comments as 'notes-to-self,' 'prompt-affirming,' or detailing 'iterative chains of decisions' rather than explaining the code's current state or intent. Humorous examples highlighted comments that were redundant, project-specific, or even misleading, suggesting Claude's comments were for other LLMs rather than human developers. While some defended them as providing historical context, the overwhelming sentiment was that these comments added unnecessary noise to the codebase, leading users to prefer Codex's leaner output.
The Ever-Evolving Ecosystem of AI Assistants
The discussion quickly broadened beyond just Codex and Claude, illustrating the dynamic and rapidly maturing landscape of AI coding assistants. Commenters detailed their experiences with a multitude of other models and harnesses, such as Sol, Fable, Luna, Gemini 3.7, Kimi K3, and Grok 4.6, often highlighting their specific strengths (e.g., speed, cost, specific task suitability). This led to observations about the 'completely contradictory anecdotal experience' across users, emphasizing that the 'best' tool is highly personal and task-dependent. Many expressed a desire for platform-agnostic solutions that allow easy model swapping, indicating a search for flexibility and optimal performance in a quickly changing field.