Codex vs Claude for Web Design: One Brief, Eight Pages

Codex vs Claude for web design, tested: one landing-page brief through GPT-6.1 Sol and Claude Opus 5.5, with screenshots, costs and what broke.

David Zhang
David Zhang

Founder, Bento

18 min read
Codex vs Claude for Web Design: One Brief, Eight Pages

We gave one landing-page brief to Codex and Claude Code, and to the models behind them, twice each, on November 5–6, 2026. Neither won outright. Codex made tighter, cheaper pages that stuck to the brief but used small text. Claude Code made more readable pages, checked its own screenshots, and invented a claim.

Questions this page answers

  • Is Codex or Claude better for web design?
  • How did we test them on the same brief?
  • What did each landing page look like?
  • What did Codex and Claude Code actually do while they worked?
  • Which one stuck to the brief, and which was easier to read?
  • What did each cost, and what would we fix?

Want both pages without two subscriptions?

We ran this whole test in one Bento workspace: both tools, one account, one bill, every page and screenshot saved. Bento can also put the page you pick on a live link with no hosting to set up. Plans start at $20 a month. Start with $5 of free credits or see Bento's plans.

What we tested: the Codex CLI and Claude Code, each in an empty folder with the brief, plus the two models on their own with one prompt in and one HTML file out. Eight pages in all.

Is Codex or Claude better for web design?

On this brief, each won different things, and the agent tools behaved like their models. Codex was the better pick for a polished page that does exactly what the brief says, cheaply. Claude Code was the better pick for comfortable reading and a stronger sense of place, if you will check what it adds.

Codex (GPT-6.1 Sol)Claude Code (Opus 5.5)
LookLight, editorial: cream, deep red or green, drawn loavesDark night sky, amber light, a moon over the loaf
Sticking to the briefEvery fact, nothing a bakery would have to stand behindEvery fact, plus "Long fermentation" in one agent run and one model run
Body text size12–13 px (down to 10 px) in all four pages16–17 px in all four pages
Accessibility scan1 contrast failure in four pages1 contrast failure in four pages
Order form after "submit"Said clearly that no order was placed, every timeSaid the order "is in" or "noted"; one agent page's form broke
Checked its own workCode checks only; never looked at the pageTook screenshots and looked at them; fixed one layout bug
Cost per page$0.09–$0.11 (model), about $0.14 (agent)About $0.25–$0.39 (model), $0.57–$0.62 (agent)
Time per agent run1 m 54 s – 1 m 59 s2 m 32 s – 2 m 54 s

Pick Codex if you will hand the page to a developer to wire up, need it to match a spec, or are generating many variations. Plan to raise the text size.

Pick Claude Code if the page has to feel like something on the first screen, and you will read every line and click every button before it ships.

How did we test them on the same brief?

We wrote one brief for a made-up business, Second Shift Bakery in Portland, and gave it word for word to four setups, two times each. Every number below comes from those eight runs.

The brief asked for one self-contained HTML file: a hero with "Pre-order for tomorrow", a three-step "how it works", a menu of three items with prices, two quotes, three FAQ answers, a pre-order form with no backend, and pickup details. It set rules too: no outside fonts, images or scripts, works on a phone and a desktop, accessible, and no new facts, prices or reviews.

SettingCodex sideClaude side
Modelgpt-6.1-solclaude-opus-5.5
Why this modelOpenAI's Codex models page tells Plus, Pro and business users to choose itAnthropic's Claude Code docs make it the default on Pro, Max and Team
Model-only runs (Nov 5)2, one prompt, 32,000-token limit, default settings2, same
Agent runs (Nov 6)2 in Codex CLI 0.164.02 in Claude Code 2.1.318
Agent settingsDefault reasoning; could write files in its folder without askingDefault (medium) effort; file edits allowed, other actions asked first

Each agent started in a fresh folder holding only brief.md, with no settings from any other project, and got the same one-line ask: build the page and save it as index.html. Both ran through Bento's model gateway rather than a ChatGPT or Claude subscription. Claude Code asked permission nine times across its two runs; we said yes every time.

Then we opened all eight pages in Chromium at 1440 px and 390 px, took screenshots, ran the axe accessibility checker, and submitted each form. All eight had every price, time and address, labelled fields, no outside requests and no sideways scrolling on a phone. One threw a script error, covered below.

What did each landing page look like?

Each side picked one mood and kept it across all four of its pages. Codex and GPT-6.1 Sol went for morning paper; Claude Code and Opus 5.5 went for the night shift.

Above the fold on a 1440-pixel desktop: Codex's two pages are cream with red or plum type and a drawn loaf; Claude Code's two are navy night skies with a moon and an amber loaf. The four agent-built pages, unedited, as Chromium rendered them on November 6.

Codex set a big serif headline with one word in italics ("We bake while you dream."), beside a scored loaf and a round sticker. The model-only pages from the day before were close cousins.

The model-only runs above the fold: GPT-6.1 Sol's two light, cream-and-rust editorial pages beside Claude Opus 5.5's two dark night-sky pages. The four model-only pages from November 5, for comparison.

Claude Code opened on a navy sky with stars, a crescent moon and a glowing loaf. Its times were set in a typewriter font, like a shift log: "21:00 Pre-orders close · Overnight · 07:00 Doors open". Its first run added a little order slip to the hero, and said in its summary that the slip was its one invented detail.

Claude Code run 1's hero: "Baked overnight. Ready when you wake up." over a navy sky, with a loaf, a moon and a paper slip reading "Order #0700, Ordered 8:52 PM, Baked overnight, Ready 7:00 AM".

The menus show the gap in detail. Codex drew each item more carefully, down to four separate swirled buns for "box of 4", and kept the prices small. Claude Code used simpler drawings on rounded cards, with larger prices and more space.

Codex run 1's menu, "The morning line-up.": a scored loaf, four swirled buns and a fruit galette, with small red prices of $9, $14 and $28.

Claude Code run 2's menu under the heading "Short list. Long fermentation.": three rounded cards with a simple loaf, two buns and a galette, prices $9, $14 and $28.

On a phone, all eight pages stacked into one column without sideways scrolling.

All four agent-built pages on a 390-pixel phone, top of the page. Claude Code's body text is visibly larger.

What did Codex and Claude Code actually do while they worked?

Both read the brief, wrote the whole page in one go, then checked it. The checking was where they differed: Codex checked the code, Claude Code looked at the page.

Codex CLI mid-run: it read brief.md, said it would use a warm cream and bakery-red palette and a demo form, then wrote index.html in one command. A real capture of the Codex terminal on November 6, about two minutes in.

StepCodex (both runs)Claude Code (both runs)
1Read brief.md, looked for project instructions (none)Listed the folder, read the brief
2Said its plan: "a warm cream and bakery-red palette" and a form that "will explain that it's a demo"Wrote the whole page in one file write (681 lines in run 1)
3Wrote the whole page with one commandSearched the file for outside URLs
4Ran a Python script checking IDs, links, labels and outside requestsTook headless Chrome screenshots at 390 px and 1440 px
5StoppedLooked at the screenshots, wrote a summary

Codex never saw its page. In run 1 it searched for a browser tool and found none; in run 2 it tried Playwright, got "No Playwright installed", and settled for the code check.

Claude Code asking for permission to read the screenshot it had just taken of its own page at 1440 pixels wide.

In run 2 the screenshots paid off. Claude Code spotted that the big decorative quote mark overlapped the review text on a phone, added padding, and took a new screenshot to confirm the fix.

Claude Code run 2 editing one line of CSS, adding right padding to the quote block, after seeing the overlap in its own screenshot.

Screenshots only show how a page looks, though. In run 1, Claude Code's summary said the form "shows an on-page confirmation". It doesn't. The script fails as the page loads ("el.addEventListener is not a function"), because a field named item clashes with a name the browser already uses. A filled-in form just reloads the page, with no message. Only clicking the button catches that.

Claude Code run 1's closing summary, listing the look, the sections, the accessibility work and the form's confirmation message, and flagging the order slip as its one invented detail.

Agent runTime (tool's own count)Output tokensCost
Codex 11 m 59 s8,628about $0.14
Codex 21 m 54 s8,636about $0.15
Claude Code 12 m 32 s17.4k$0.57
Claude Code 22 m 54 s19.3k$0.62

Claude Code's costs are what it reported in its own /cost screen. Codex doesn't show a cost, so its figure is its token counts multiplied by the gateway's list price. Claude Code's times include a few seconds waiting for each yes.

Which one stuck to the brief?

Codex did. Every page used every fact, but only the Claude side added ones a bakery would have to stand behind, and the agent loop didn't stop it.

Claude Code's second run headed the menu "Short list. Long fermentation." and added "Three things, made properly, every night." Opus 5.5 on its own had written "Short menu. Long ferment." the day before. The brief says nothing about fermentation. Then Claude Code's summary said the page used "only the facts, prices and quotes it gave".

Its first run invented the order slip in the hero, and said so.

Codex's additions were connective copy, like "A little something to look forward to, right here in the neighbourhood" and "Sharing optional." Nothing a customer could hold the bakery to.

A Reddit thread that Google summarises in its results says the same: Claude fills in missing UI details, Codex is more literal. Filling in helps when the brief is thin, and hurts when it isn't.

Which page was easier to read and use?

Claude's, on every page. Its body text was 16–17 px, a comfortable size. Codex and GPT-6.1 Sol set most paragraphs at 12–13 px, some at 10 px. The agent loop didn't fix it: Codex never looked at its page.

The "how it works" steps from Codex run 1 on a 390-pixel phone: large serif headings, with the step text underneath set at about 10 pixels. Shown at twice phone size so the type is legible here. On the phone it is half this.

DesktopTypical paragraphSmallest paragraphAccessibility issues (axe)
GPT-6.1 Sol, model runs 1 / 212 / 13 px11 / 11 px0 / 0
Codex, agent runs 1 / 212 / 12 px10 / 10 px0 / 1
Opus 5.5, model runs 1 / 216 / 17 px12.5 / 14.4 px1 / 0
Claude Code, agent runs 1 / 217 / 17 px12.5 / 12.8 px0 / 0

The two contrast failures were small. Opus 5.5's first page put a big decorative "03" on amber at 1.75 to 1, where large text needs 3 to 1. Codex's second page set its step numbers in a soft brown at 4.24 to 1, just under the 4.5 to 1 that normal-size text needs.

The forms showed the biggest difference in judgement. None had a backend, as the brief said. All four Codex and GPT-6.1 Sol pages said so after a submit, for example "This is a demo, so your order has not been sent or reserved." The Claude side's pages didn't.

After a filled-in submit: Codex's form says "This is a demo, so no order has been placed." Claude Code's says "Thanks, Test! Your pre-order for Country sourdough loaf on Tuesday is in."

Opus 5.5's pages said the order "is noted"; Claude Code's second page said it "is in"; its first showed nothing, because of the script error. Ship either by mistake and a customer turns up for bread nobody ordered.

Opus 5.5's model-only form after a submit, with a green box reading "Thanks, Test! Your pre-order … is noted for Tuesday".

How much did each page cost, and how long did it take?

The Codex side cost a quarter to a third as much per page. The agent loop added about half again for Codex and nearly doubled it for Claude Code.

SetupCost per pageTime per pageOutput tokens
GPT-6.1 Sol, model only$0.093 and $0.110110 s and 130 s9,190 and 10,915
Codex agentabout $0.14 and $0.15119 s and 114 s8,628 and 8,636
Opus 5.5, model onlyabout $0.39 and $0.25159 s and 99 s19,453 and 12,515
Claude Code agent$0.57 and $0.62152 s and 174 s17.4k and 19.3k

GPT-6.1 Sol's model costs are what the gateway billed, and Claude Code's are its own /cost figures. The rest are token counts multiplied by the gateway's list prices. Times include network delays on two afternoons, so treat them as a rough sense, not a benchmark.

Cents don't matter for one page. They matter when an agent rebuilds the page every time you ask for a change. On a Bento plan those runs come out of one monthly pool of credits instead of two separate bills, and you can top up if a redesign week runs long.

What did other side-by-side tests find?

Two recent public tests ran both tools on the same prompt, and both found Codex's first impression stronger.

  • MindStudio's write-up (August 15, 2026) put one prompt for a Typeform-style app through Codex and Claude Code. Codex's build "looked more polished on the surface", with a stronger landing page; Claude Code's was plainer but "far more usable".
  • Nate Herk's video (August 26, 2026) built eight websites with the same prompts and brand guidelines. In his description, Codex won most of the design rounds using about a third of the time and a fifth of the output tokens, and detailed prompts made the two tools' results nearly identical.

Our result agrees on polish and cost, and adds what a screenshot can't show: type size, contrast, and what the form tells a customer.

What would we fix before shipping either page?

None of the eight was ready to ship as delivered. Each needed a short list of edits.

Codex's page

  1. Raise body text to at least 16 px and the smallest labels to 12 px or more.
  2. Darken the step numbers in run 2 to pass contrast.
  3. Connect the form to a real order system and replace the "demo" message. Bento can build the part that saves the order, so this step doesn't need a separate backend.

Claude Code's page

  1. Delete "Long fermentation" and the order slip unless the bakery confirms them.
  2. In run 1, rename the item field so the form's script runs at all.
  3. Rewrite the confirmation so it only says "is in" once an order is actually saved.
  4. Cut a section or two; at 4,100–4,700 px on desktop the pickup details sit a long scroll from the top.

Whichever tool you use, click every button yourself. Claude Code's screenshots caught a layout bug and missed a broken form.

When the page is ready, our guide shows how to take an AI-built page live. In Bento, apps and sites go live at their own link.

Does this tell you which coding tool to buy?

For a one-page design job, mostly. But we gave each tool a blank folder, one ask and no follow-up. In a real project you'd ask for changes, and that back-and-forth is where tools differ most. If you'd rather describe the changes than run a coding tool yourself, Bento takes the request in plain words and makes the edit.

For the rest of the decision, like usage limits, pricing and how each tool works in a real repository, see our full Codex vs Claude Code comparison and Claude Code vs Cursor vs Codex. We also didn't test each vendor's most expensive model: GPT-6 Astra or Claude Fable 5.1.

Can you run both in Bento without paying for two tools?

Yes. That's how we ran this test. One Bento workspace sent the brief to GPT-6.1 Sol and Claude Opus 5.5, ran Codex and Claude Code against both on one account and one bill, saved all eight pages, and took the screenshots you see here.

The honest answer to "which is better" keeps being "it depends on the page", so running your own brief through both and keeping the better parts beats picking one.

Bento can also put the page you pick on a live link without setting up hosting, and build the part of the form that actually saves the order.

Bento's plans start at $20 a month for Starter, with Pro at $50 and Max at $200. Every plan includes all of Bento and your whole team at no extra cost per person, and every new account starts with $5 of free credits.

Make anything. One subscription. See Bento's plans.

Frequently Asked Questions

Is Claude better than Codex for frontend design?

Not across the board. In our same-brief test, Claude Code and Opus 5.5 produced more readable pages with more atmosphere. Codex and GPT-6.1 Sol produced more polished, on-brief pages for a quarter to a third of the cost.

Which is cheaper for UI work, Codex or Claude Code?

Per page in our test, the Codex agent cost about $0.14–$0.15 and Claude Code $0.57–$0.62, both at gateway prices. Used without the agent, GPT-6.1 Sol cost $0.09–$0.11 a page and Claude Opus 5.5 about $0.25–$0.39.

Does Claude Code check its own design work?

In our runs, yes: it took headless Chrome screenshots at phone and desktop widths and looked at them, and once fixed an overlap it saw. It didn't click anything, so it missed a form whose script failed to load.

Which model does Codex use?

OpenAI's Codex models page tells users on Plus, Pro, Business, Enterprise and Edu to choose GPT-6.1 Sol. GPT-6 Astra is the more capable and more expensive option. Free and Go plans are pointed to GPT-6 Luna.

Which model does Claude Code use?

Claude Code's default is Claude Opus 5.5 on Pro, Max, Ultra, Team, Enterprise and the Anthropic API, according to Anthropic's model configuration docs. You can switch to Sonnet 5.5 or Fable 5.1 with the /model command.

Can I reproduce this test?

Yes. Put the brief in an empty folder, ask each tool to build it as index.html, then open the files at 1440 px and 390 px and submit every form. Expect different pages each run; our two Claude Code runs differed by about 2,000 output tokens and 20 seconds.

Crea lo que sea. Una suscripción.

Posts, reels, anuncios, apps y juegos, hechos con la mejor AI.

Créditos gratis para probarlo todo. Nada que configurar.

Artículos relacionados

Build and Deploy a Web App Using Only AI
AI & Automation14 min read

Build and Deploy a Web App Using Only AI

No dev team? No problem. Build and ship a fully functional web app using only AI — from idea to deployed in days.

Bento Team
Bento Team