Codex vs Claude Code on a Product Page and a Mobile Revision

The two revised product pages at 390 pixels wide, Claude Code on the left and Codex on the right
On this page

We gave Claude Code and OpenAI Codex the same perfume packshot, asked each for a product page, then requested one mobile revision. Both delivered all four required elements: the photograph, the product name, the supplied line and a working link to the bottle details. The two jobs on Claude Code used 47 Flowith credits together. The two on Codex used 8. Open the pages below to decide how much you value the differences in layout.

What the two runs cost

On November 20, 2026 we used the default model offered for each coding agent on the same Flowith account and canvas. Each agent started with a fresh card. Its revision continued from its own first result. We recorded the final billed meter and rounded the interval from the recorded start to delivery to the nearest second.

Agent and modelTaskTokens reportedBilled creditsElapsedDelivered page
Claude Code · Claude Opus 5.5First page32,22229108 sOpen first version
Codex · GPT-6.1 SolFirst page13,7566118 sOpen first version
Claude Code · Claude Opus 5.5Mobile revision11,8691883 sOpen revised version
Codex · GPT-6.1 SolMobile revision1,440250 sOpen revised version

Claude Code delivered the first page about ten seconds sooner. Codex delivered the revision about 33 seconds sooner. A different brief, model, cache state or account rate can change the bill. These figures cover two versions per agent and exclude the subscription fee. Check current coding rates on Flowith.

Run the same brief

Attach the original perfume photograph and paste the following text into a fresh card for each agent.

Build a static English product landing page for the supplied Orange Sanguine perfume packshot. Use the supplied image as the hero, with alt text. Include the exact heading “Orange Sanguine”, the line “A study in citrus”, and a “View the bottle” link to a detail section on the same page. The detail section should describe only what is visible in the photo. No prices, checkout or external dependencies. Make a responsive static HTML site, with the asset included, and deliver a viewable web page.

Open each result at 390 pixels wide. In both first versions the headline appeared above the photo. On the Claude Code page the “View the bottle” link measured 51.25 pixels high. On the Codex page it measured 39 pixels. Give each agent this follow-up on its own card:

Revise the same page. At 390 pixels wide put the product photo before the headline, show the complete blue cap and bottle, and make the “View the bottle” link at least 44 pixels high. Preserve the desktop layout and all original copy. Deliver the revised viewable web page.

What changed at 390 pixels

In both revised pages the whole blue cap and the base of the bottle fit in the frame, with the photo above the headline. Both linked to an existing detail section and had no horizontal overflow. The visible words matched each agent’s first version after normalizing whitespace.

Claude Code’s revised photo starts about 63 pixels from the top of the viewport. Its compact headline leaves room for the brief’s call to action below the product. Codex starts its photo at 121 pixels and breaks the larger product name across two lines. The revised links measure 46.13 pixels and 45 pixels high respectively. Both clear the requested 44-pixel minimum.

What stayed the same on desktop

At 1280 pixels wide, the hero photo kept its original placement and size in both revisions. Codex also kept the original desktop link height of 39 pixels. Claude Code changed that link from 51.25 to 46.13 pixels, although the requested adjustment specified mobile. If desktop fidelity matters in a handoff, keep the first version open beside the revision and measure the affected element at both widths.

For this small static page, both agents delivered a usable revision on the first follow-up. Codex used fewer credits and kept the desktop button height. Claude Code used more compact mobile typography and put the product higher in the viewport. Readers evaluating a larger codebase can reuse the method: write down the viewport, exact copy, target element and expected behavior before either run, then open the actual deliverables. Compare Codex access and billing options.