In one portfolio build, Dibakar Ghosh preferred Claude Code’s more distinctive design and Codex’s more personal copy; he also reported that Codex finished sooner. His comparison used one run per tool, with different underlying models, so the result offers a task-specific preference rather than a general ranking. Ghosh’s account was published October 12, 2026.
What the portfolio test compared
Ghosh gave both agents the same portfolio prompt and résumé, and permission to research him online. The assignment asked each to research the subject, ask clarifying questions, make design and technical decisions, and build a site. Claude Code ran as v2.1.295 with Sonnet 5.5 at xHigh effort; Codex ran as v0.162.0 with GPT-5.6 Terra at xHigh effort.
The underlying models differed, so the output differences cannot be attributed to the coding agents alone.
How the two portfolio results differed
| Comparison point | Claude Code | Codex |
| Configuration in this run | v2.1.295; Sonnet 5.5; xHigh effort | v0.162.0; GPT-5.6 Terra; xHigh effort |
| Design | Ghosh preferred its more distinctive, editorial-style design | Ghosh described the layout as more conservative and corporate-looking |
| Copy | Ghosh found it generic; the homelab section emphasized specifications over projects | Ghosh found it more personal |
| Reported completion time | About 30 minutes | Just over 20 minutes |
Both agents needed follow-up prompts. Ghosh also planned to revise Codex’s copy himself, despite preferring it in this run.
Match the choice to the work you want help with
For Ghosh, design was harder to direct than copy was to revise, so he chose Claude Code for his own needs. A designer or developer who wants more help with writing might instead favor Codex, he said. That’s a useful way to think about the split: which part of your portfolio would you rather refine yourself—its visual direction or its words?
Codex finished sooner in this run
Ghosh reported just over 20 minutes for Codex and about 30 minutes for Claude Code. Those are approximate timings from one run each, not a general speed comparison; both builds also needed follow-up prompts.
A shared change log can pass work between agents
Claude Code and Codex do not have to be treated as an either-or choice. In a separate calculator demonstration, Wyndo and Dheeraj Sharma used a shared changes.log to record completed work, affected files, and context for the next agent. Claude Code made an initial change, Codex continued and checked the interface, and Claude Code then resumed. The One Shot Show demonstration offers one concrete handoff pattern for keeping project context visible as agents take turns.
