Writing an Essay with AI: Codex vs Claude Code
The author compares the writing quality of OpenAI Codex (GPT 5.2/5.3) and Anthropic's Claude Code (Opus 4.5/4.6) in the process of writing a 16-chapter essay called 'Science Catch-Up'. The essay critiques scientism and proposes a framework for evaluating knowledge.
Why it matters
This comparison highlights the rapid progress in language model capabilities for creative and analytical writing tasks, with Anthropic's Claude Code outperforming OpenAI's Codex in a demanding essay project.
Key Points
- 1Codex (GPT) produced verbose, repetitive writing with rigid formatting and a conciliatory tone
- 2Claude Code (Opus) provided more natural prose, better adherence to editorial guidelines, and stronger research and referencing capabilities
- 3Claude Code was able to generate programmatic figures and diagrams for the essay
Details
The author used both Codex and Claude Code to write the 'Science Catch-Up' essay, a complex work that critically examines the limits of the scientific method. The initial drafts written with Codex (GPT 5.2/5.3) suffered from issues like repetitive expressions, rigid 'example' formatting, and a bland, conciliatory tone that did not match the essay's combative and provocative nature. In contrast, the later work done with Claude Code (Opus 4.5/4.6) showed immediate improvement in terms of natural language, adherence to editorial guidelines, and research capabilities. Claude Code was also able to generate programmatic figures and diagrams for the essay. However, the author notes that Claude Code still struggles with maintaining narrative context when using parallel subagents, and the output still requires heavy review and revision.
No comments yet
Be the first to comment