For WordPress development, start by comparing Claude Code and OpenAI Codex on a small change in your actual development environment. Add a Gemini-based workflow if it fits the tools you already use. Choose the one that produces a maintainable result with the least correction, including the time needed to test it.
A useful answer needs more detail than “Claude writes better code” or “ChatGPT is faster”. WordPress work includes PHP, browser behaviour, editable content, plugin compatibility and deployment. An assistant can handle one part well while leaving another unfinished.
This comparison draws on published tests and official product documentation. The WordPress exercise below is a reproducible test you can run; it is not presented as a completed three-model benchmark.
Compare the model and its working environment
Claude, GPT and Gemini are model families. Claude Code, Codex and an editor or command-line integration give those models ways to inspect files, make changes and use tools. The result depends on both.
Codex CLI can work with a local codebase. Claude Code also works through a cycle of inspecting, acting and checking. Google's Gemini CLI documentation describes file, shell and project-context tools. The capabilities enabled in a particular installation still need checking.
Giving one tool a complete repository and another a pasted screenshot tests different workflows. That can be useful when choosing how to work, but it does not isolate model quality.
| Your work | What to compare | What a useful result looks like |
|---|---|---|
| Fix an existing theme | Repository access and a focused code change | The bug disappears and neighbouring pages still work. |
| Build an editable landing page | Access to the actual editor or content model | The owner can change text, images and links after handover. |
| Extend a plugin | PHP implementation and WordPress conventions | Inputs, permissions, output and failure paths are handled. |
| Improve a mobile layout | Browser inspection and visual judgement | Real content fits, controls work and the desktop layout survives. |
| Explain an unfamiliar project | File inspection and clear reasoning | Explanations point to the relevant code and identify uncertainty. |
For a small explanation, a chat interface may be enough. For changes across a theme, a workflow that can inspect the project and verify the result is a better starting point.
What published comparisons actually show
Stefan Wirth's WordPress and Elementor comparison reports testing Claude Opus 5.5 in Claude Code against GPT-6 Astra in Codex through the same Elementor MCP. He reports 33 minutes for Claude and 18 for Codex across the initial two rounds, with more reusable styling in Claude's first result. His revised brief changed the output. This is one author's Elementor experiment, not a general WordPress ranking.
In an April 2026 five-task comparison, J. D. Hodges tested Claude Opus 4.6, GPT-5.3-Codex and Gemini 2.5 Flash alongside local models. He preferred Claude's implementation quality and Codex's code review in that sample. Those older versions and mixed interfaces cannot establish today's winner.
AIMultiple's coding benchmark illustrates another complication: its CLI group used a shared Claude model to compare orchestration, while its editor group used a different model. A tool's name in a chart does not necessarily identify the provider of the model doing the work.
The useful lesson is to inspect the task, model, interface and acceptance criteria before adopting a result. A screenshot of an attractive page says little about whether a form stores enquiries or an editor can change its content.
A WordPress task that reveals practical differences
Use a disposable local WordPress installation with the same theme, plugins and sample content for each run. Give every candidate a fresh copy. Keep production credentials and real enquiries out of the fixture.
The task is a service directory for a fictional bicycle workshop. It is small enough to review and still tests several parts of real work: native editing, a custom output component, localisation, price handling and responsive layout.
Give each tool this brief:
Build a small WordPress plugin for a fictional bicycle workshop.
Create an editable service content type with a title, description,
price in integer cents, and duration in whole minutes. Use EUR.
Editors must be able to manage services; subscribers must not.
Provide [workshop_services] to display published services.
Show a zero price as Free. Hide missing prices and durations.
Do not show drafts or invent missing information.
Use WordPress APIs. Protect administrative saves with appropriate
permission and nonce checks. Validate input and escape output.
Make visitor-facing strings translatable.
Use ordinary HTML and CSS with no external service or new library.
The cards must fit at 390 px and 1440 px. Show an understandable
empty state when no published services exist.
Inspect the supplied environment first. Implement the plugin,
explain how to install it locally, and report what you verified.
Do not deploy or alter unrelated site settings.
Keep that prompt identical. Record the precise model, tool version, settings, date and starting environment separately. If one candidate cannot access a tool the others use, record that difference rather than silently compensating for it.
Check behaviour before judging the explanation
Create three published services: a paid repair, a free inspection and a service without a price. Add a draft service that must stay private. Then change the data in the editor and reload the public page.
| Check | Pass condition |
|---|---|
| Native editing | A normal editor can update the service without changing source code. |
| Published content | Public output includes published services and excludes the draft. |
| Zero and missing values | Free and unknown prices remain distinct. |
| Saving | An authorised edit persists; an unauthorised request cannot change the data. |
| Output handling | Test punctuation and HTML-like text display safely. |
| Layout | Cards fit at the two stated widths without page-wide scrolling. |
| Empty state | Removing the published fixtures produces the specified message. |
| Scope | Other pages, settings and existing content remain unchanged. |
WordPress distinguishes permission checks, validation and output escaping. Its security guidance is a useful review reference. A nonce helps with request integrity; it does not replace authorisation. The nonce documentation makes that distinction explicit.
Inspect the resulting code as well as the visible page. A working demonstration can still contain overly broad changes or an unsafe save handler.
Give each candidate the same correction opportunity
Preserve the first output. Run the acceptance checks and provide a factual failure list, such as “the draft service is visible” or “the zero price is hidden”. Allow the same number of correction rounds and the same time budget for each candidate.
Record first-pass behaviour and corrected behaviour separately. A tool that reaches a good result after several interventions may still suit your work, but those interventions belong in the comparison.
Repeat the exercise with a fresh session before drawing a strong conclusion. If the outcome changes, report the variation. Do not average away a serious permission failure just because the CSS looks good.
Compare the cost of a completed change
The useful cost includes your review time. Suppose one workflow takes 15 minutes to generate a plugin and another takes 25. If the first needs an hour of repair and the second needs ten minutes, generation speed alone gives the wrong purchasing signal. These are illustrative numbers, not measured results.
Separate subscription allocation, metered API spend and any paid tools. Keep usage estimates distinct from an invoice. Record how much context you supplied and whether retries or additional agents were involved.
The broader AI-model comparison helps build a shortlist. This exercise answers a narrower question: which available workflow can finish your WordPress task to an acceptable standard?
How I would make the choice
For repository changes, I would trial Claude Code and Codex first, using the same local fixture. I would include Gemini when it is already part of the team's workflow or when a third candidate could resolve a specific weakness in the first two. This is a testing order, not a measured league table.
For an owner who needs an editable website, the handover test matters most: change a service, replace an image, submit a form and restore a backup. For a developer, add code review and regression checks. Keep the tool that completes those tasks reliably at a cost you can justify.
If your project already has a working theme, begin with a bounded improvement. A model does not need to rebuild the entire website to prove useful.
Putting AI to work on your website?
Choosing a model is one part of a working feature. If your business needs a support assistant, content workflow or connection to an existing system, I can help define and build the interface, integrations and checks it needs.