Which AI is better at design work?
On design work that goes back and forth, the model that used fewer tokens per round cost about a third as much.
What the reel showed
One brief for an interactive 3D lightsaber in a single web page: built from basic shapes, a glowing blade, parts that slide out when you click them, an explode button and a colour picker. Claude Fable 5.1 ran in Claude Code, GPT-6 Astra in my own agent.
Each went through three rounds of art direction from me. The reel shows the third version from each.
Fable came to $19.29 across the three rounds and Astra to $6.50. Both models have the same price per token, so the whole difference is how many tokens each used to get there, and Astra used far fewer.
| Entrant | Tool | Cost | Verdict |
|---|---|---|---|
| Claude Fable 5.1 | Claude Code | $19.29 | Three rounds, checked in a browser |
| GPT-6 Astra | Hermes agent | $6.50 | Three rounds, far fewer tokens |
Both got the same brief and three rounds of art direction each; costs are read off each run's own bill and can differ from the figure shown in the reel.
What this means for your business
Design work is rarely one shot. You ask for a landing page, then change the colours, then the headline, then the layout. Every round costs tokens, so a model that is efficient per round gets cheaper with every change you make.
Uses: landing page drafts, a client report layout, a pitch page, a mockup of a new product page before you brief a designer, a one-off visual for a sales call.
When someone shows you what a model built for a few dollars, ask how many rounds that took. The number under a finished result often covers three or four.
What I’d use
For design passes and mockups I lean on Astra now. It cost about a third as much here. For long jobs inside a real codebase I still use Claude, because it checks its own work in a browser, and that is where Fable spent many of its extra tokens.
Either way, run your first real design brief through both and compare the result before you settle on one. I do this on my subscriptions: Claude Code on my Claude plan, Codex CLI on my ChatGPT plan.
Try it yourself
Design one real page for your business with both models, give each the same two rounds of feedback, and compare what you get. Use the subscriptions you may already have, or OpenRouter if you want an exact cost comparison like the reel's.
Route A: Claude Code on your Claude plan
This is how I work every day. Claude Code is included with Claude Pro and Max, so runs draw on your plan's usage limits instead of a bill per call, and you can go back and forth as often as you like. Fable is included on Max; on Pro it runs on paid usage credits.
Route A: Codex CLI on your ChatGPT plan
Codex CLI is OpenAI's coding agent. Run it and choose Sign in with ChatGPT to use it on a ChatGPT Plus or Pro plan. GPT-6 Astra is available in Codex on both.
Route B: OpenRouter, for the exact bill
A subscription shows up as usage against your plan, not as dollars per run. To compare the bill the way the reel did, use one OpenRouter account and add Claude Fable 5.1 and GPT-6 Astra in the chat. Both list at $10 per million input tokens and $50 per million output tokens there, so any difference in your bill comes from how many tokens each one uses.
Brief a real page
Fill in the brackets. Keep the brief the same for both. On Route A, start each agent in its own empty folder.
Design a landing page for [your business, one line] as a single HTML file called index.html. Audience: [who buys from you] Goal: get them to [book a call / request a quote / buy] Include: a headline and subline, three benefits, a testimonial placeholder, a "how it works" section, and one clear button. Style: [two or three words, e.g. "calm, premium, lots of white space"]. It must look good on a phone. If you can write files, save it in this folder. Otherwise return the whole file in one code block.
Give the same feedback twice
Open each page in your browser. From the chat, save each answer as a .html file first. Then send both the same two rounds of notes, for example 'shorter headline, make the button stand out', then 'add a section for common questions'.
Compare the page and the cost
Pick the page you would actually publish. On Route B, open your OpenRouter activity page and compare what each model cost across all three rounds.
Something I haven't tried yet, but you can
Higgsfield, an AI image and video platform, has an official MCP connector for AI agents. In Claude Code you can ask the agent to set up Higgsfield for you, and Codex CLI can add it as an MCP server. Then the agent that designed your page could also generate the hero image, product shots and ad variations to match it. It uses a paid Higgsfield plan and its credits. I have not tested this, so treat it as an idea to try.
Links
- Claude CodeAnthropic's coding agentclaude.com
- Claude Code with Pro or MaxRun it on your Claude subscriptionsupport.claude.com
- Codex CLIOpenAI's coding agent, signs in with a ChatGPT plangithub.com
- GPT-6 AstraOpenAI's announcementopenai.com
- OpenRouter chatSend one prompt to several models side by sideopenrouter.ai
- GPT-6 Astra on OpenRouteropenrouter.ai
- Claude Fable 5.1 on OpenRouteropenrouter.ai
- Higgsfield MCPHiggsfield's official connector for AI agentshiggsfield.ai
Want a landing page or client report designed with AI, properly?
The drafts are cheap now. Turning one into a page that matches your brand and actually converts is the part I do.