Claude Opus 4.6 vs. GPT-5.3 Codex: How I shipped 93,000 lines of code in 5 days

What It Covers

Claire Vaux tests OpenAI's GPT-5.3 Codex and Anthropic's Claude Opus 4.6 models side-by-side on real coding tasks, shipping 93,000 lines of code across 44 pull requests in five days. She evaluates their strengths for redesigning marketing sites and refactoring complex codebases, revealing distinct use cases for each model.

Key Questions Answered

• •Model pairing workflow: Use Claude Opus 4.6 to build new features and creative implementations at 80-90% completion, then switch to GPT-5.3 Codex for architectural review and edge case detection. Codex identifies issues Opus missed, Opus quickly implements fixes, creating a principal engineer plus product engineer dynamic that ships production-ready code faster than either model alone.
• •Codex literal interpretation problem: GPT-5.3 Codex overfits to exact prompt wording, creating unintended results like making a homepage headline say "dense product workflow" when asked for "content-dense site design." The model follows instructions precisely but lacks nuanced interpretation for creative tasks, requiring multiple correction rounds where each prompt causes overcorrection in the opposite direction.
• •Git-first development interface: Codex desktop app surfaces Git primitives as first-class features including branches, work trees for parallel agent work, visual diff panels showing line-by-line changes, and pull request creation. This approach teaches non-technical users version control concepts while enabling advanced users to run multiple agents simultaneously on separate work trees without conflicts.
• •Opus greenfield superiority: Claude Opus 4.6 excels at broad, creative tasks like complete site redesigns, independently planning multi-step implementations and maintaining consistent design systems across multiple pages. It shipped a production-ready marketing site redesign matching brand aesthetics after one design correction round, versus Codex which only completed two pages despite similar prompting and required constant guidance.
• •Production velocity metrics: Shipping 44 pull requests with 98 commits across 1,088 files, adding 92,000 lines and removing 87,000 lines in five days demonstrates ROI of premium AI coding models. This included five MCP integrations, component refactors, and vector store replatforming that would require months with traditional development, justifying Opus 4.6 Fast's cost at 150 dollars per million output tokens.