SkillArena tests UI skills for AI coding agents (SKILL.md files like frontend-design, better-typography, fixing-accessibility). Each skill gets the same five briefs: a SaaS landing page, a signup form, a dashboard, a mobile app screen, and a page for SkillArena itself. A "no skill" row uses the same prompt without any skill, so you can see what the skill actually changes.
Until now every page came from Claude. This week I added GPT for the first 22 skills plus the no-skill row, and put them in the matrix as a second row under each skill.
Setup
| Claude | GPT | |
|---|---|---|
| Model | claude-opus-5-5 | gpt-6.1-sol (reasoning: medium) |
| How | API, one call per page | Codex (codex exec), a fresh session per page in an empty read-only folder |
| Runs | 3 per skill × brief | 1 per skill × brief |
| Pages | 345 | 115 |
| Failed pages | 0 | 0 |
Same briefs, same skill text, same output rule (one self-contained HTML file). The pages are screenshotted the same way and served sandboxed.
The setups are not identical, and that matters. Codex adds its own agent instructions on top of mine, and GPT has one run per pair, so a single GPT page can be a lucky or unlucky draw. Compare GPT with GPT's own no-skill row, not just with Claude.
1. Each model has a default font
Counting the fonts each page loads:
- Claude: Inter in 190 of 345 pages (55%), then Fraunces (68), JetBrains Mono (62), Bricolage Grotesque (37).
- GPT: DM Sans in 73 of 115 pages (63%), then Manrope (47), Space Grotesk (27). Inter in only 7.
So "AI sites all use Inter" is really "Claude sites use Inter". GPT has its own house font. A skill that says "avoid generic fonts" is pushing against a different default on each model.
2. GPT writes shorter landing pages
Average full-page height on desktop:
| Brief | Claude | GPT |
|---|---|---|
| SaaS landing page | 4,428 px | 3,075 px |
| SkillArena home | 3,934 px | 2,167 px |
| Signup form | 1,368 px | 1,105 px |
| Dashboard | 1,286 px | 1,323 px |
On the open-ended briefs Claude keeps adding sections and GPT stops earlier. On the tightly scoped briefs (dashboard, form) the two are close. The HTML files are about the same size (27 KB vs 29 KB on average), so GPT's pages are denser rather than smaller.
You can see it on the SkillArena brief. Claude spends the first screen explaining the idea. GPT puts the matrix itself above the fold.

no-ai-design-slop skill. Claude · GPT
On a scoped brief like the dashboard, the two land close: KPI cards, a line chart, sales by location.

No skill, dashboard. Claude · GPT
3. The no-skill pages look different
With no skill at all, the SaaS landing page brief gave:
- Claude: a centred hero, an indigo/violet accent, a gradient word in the headline, a "New" pill above it. The look people call "AI slop".
- GPT: a left-aligned hero, very large tight-set type, a muted green palette, a calendar mock on the right.

No skill, same brief. Claude · GPT
Add a skill and both pages change, but not in the same direction. With Anthropic's frontend-design skill, Claude left the light purple look for a dark navy page with a yellow button. GPT went the other way, to a light lavender background with an indigo button: closer to Claude's no-skill default than to GPT's own.

Same skill, same brief. Claude · GPT
Gradients follow the same pattern: Claude pages average 2.3 gradient declarations, GPT pages 0.8.
4. GPT starts with more accessibility basics, so skills add less
Counting a few easy-to-grep signals in the HTML:
| Claude, no skill | Claude, with skills | GPT, no skill | GPT, with skills | |
|---|---|---|---|---|
Handles prefers-reduced-motion
|
40% | 85% | 80% | 91% |
aria-* attributes per page |
15 | 26 | 20 | 29 |
Skills move Claude a lot; GPT already does most of this without being asked. If you use GPT, an accessibility skill has less to fix. These are grep counts, not an audit: they show a page tried, not that it is accessible.
5. What I'd take from this
- A skill's effect depends on the model. The same SKILL.md lands on a different starting point. A skill written against Claude's defaults ("don't use Inter, don't use purple gradients") targets habits GPT doesn't have.
- Judge a skill against its own model's no-skill row. That's why every GPT row on the site sits next to GPT's no-skill row.
- One run is not enough to rank anything. I wouldn't call any single skill "better on GPT" from this data, only the broad patterns above, which hold across 22 skills.
Look for yourself
Every page is clickable and runs live (sandboxed):
- The matrix, with a GPT row under each skill: https://skillarena.dev/
- Working versions of the SkillArena page that read real data, from both models: https://skillarena.dev/live/
- How it's run, including the GPT caveats: https://skillarena.dev/about/#other-models
- Code and data (MIT / CC BY 4.0): https://github.com/xyunsh/skillarena
Questions I'd like answers to: which skills would you add, and should I run GPT on every skill, 3× per pair like Claude (about 1,650 more pages)?
Top comments (0)