Fable 5.2 vs Opus 5.5 for Web Design: Costs, Tests and Defaults

Which Claude model should you use for a landing page?

Choose by the job. Anthropic's general advice is to 'start with Claude Opus 5.5 for most workloads' and to use Fable 5.2 'for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short'. The Fable 5.2 launch post makes no claim about front-end work and its prompting guide has no design section, so the design case rests on outside tests. Fable 5.2 lists at $8 per million input tokens and $40 per million output, twice Opus 5.5 at $4 and $20. Here is the routing we would use.

What did the 18-site Velox Themes test find?

Widya Bayu W gave both models three invented businesses across three rounds, using Claude Code desktop 2.1.317 at Medium effort, with one prompt per site and no follow-ups. Costs are what the work would come to at Anthropic's API prices. Read it knowing the conflict: he co-founded Velox Themes, which sells the Design.md guides used in rounds two and three and offers the MCP tool that delivers them, and he says so. The quality calls are his own. The cost and time figures are the useful part.

How much does a landing page cost on each model, in rupees?

The averages below are from the Velox table, converted at the RBI reference rate of Rs 93.8579 per US dollar on 9 November 2026 and rounded. The interesting number is the last column. With a short prompt, Fable 5.2 cost 2.4 times as much. Give both models a design guide and the gap falls to 1.3 times, and Opus 5.5 was actually the slower of the two in that round.

• •Worked example, 20 guided pages a month: ($2.97 minus $2.36) x 20 = $12.20, or about Rs 1,145 more on Fable 5.2.
• •Worked example, 20 short-prompt pages a month: ($2.18 minus $0.92) x 20 = $25.20, or about Rs 2,365 more on Fable 5.2.
• •Set that against one hour of your designer's time. At these volumes the model bill is the smaller number, so rework is the cost to manage.
• •On a subscription this work draws down plan limits instead. See /blog/claude-pro-max-ultra-usage-limits-november-2026-india.

Does the frontend-design skill really halve the cost?

Once, on one model, and the reason matters. A pull request to the which-ai gallery ran Fable 5.2 through headless Claude Code at xhigh effort, five landing pages per run from one prompt. Without skills it took 62 minutes, 387 turns and $41.13, about Rs 3,860. With Anthropic's frontend-design skill it took 43 minutes, 203 turns and $20.52, about Rs 1,926. The same report shows the opposite on Fable 5.1, where the skill run cost more: $12.29 against $4.95. Its author puts the expensive Fable 5.2 run down to self-checking, with 73 screenshots and 76 scripted browser checks. There is one run per condition and no score for how the pages look. Install the skill for the design direction it gives, and do not budget on it halving your bill. The stronger lesson is effort. Anthropic reserves xhigh for 'your hardest long-running coding and agentic work, where you've measured a gain', and a landing page rarely qualifies.

Can Fable 5.2 see its own pages better than Opus 5.5?

On chart reading, yes, by Anthropic's own vendor-reported numbers. On Chartography, with no tools at each model's highest available effort, Anthropic reports 73.4% for Fable 5.2, 64.4% for Opus 5.5 and 44.8% for Fable 5.1. At Medium effort, the Claude Code default, Fable 5.2 scores 66.4%, so its lead over Opus 5.5's top score shrinks to two points. Reading charts is also a different skill from judging whether a hero section has weak hierarchy, and we found no published benchmark for that. In the field both models run the loop: Wooldridge reports that Fable 5.2's builds 'both checked their own work in a browser before finishing', and Muzli's Petras Baukys had Opus 5.5 work for about 80 minutes through four rounds of screenshots and fixes. Pay for the sharper eyesight when a page is dense with charts or data, after testing it on your own screens.

What are Claude's five design defaults?

This is the one place Anthropic has written down what its model does unprompted. From the Opus 5.5 prompting guide: 'Asked for frontend work without design direction, Claude Opus 5.5 falls back on a few default styles, and a general instruction such as "avoid a generic AI look" mostly swaps one default for another.' Its example names five patterns. The Fable 5.2 guide has no matching section, so applying this to Fable 5.2 is our inference. The tests support it: all six round 1 Velox pages from both models were cream or warm off-white, and Wooldridge found 'Both landing pages lean on monospace labels for character'. For why models converge at all, see /blog/why-ai-websites-look-the-same-how-to-make-claude-design-better.

• •Cream or off-white background: name your actual background hex. If your brand truly is cream, keep it and say so.
• •Italic accent words in headlines: state how emphasis works instead, such as weight or colour on a whole phrase.
• •Numbered '01/02/03' section labels: keep them only where order carries meaning, such as a process.
• •Monospace labels: restrict monospace to code, prices or data where alignment helps.
• •Pill-shaped buttons: give one corner radius in pixels and use it across buttons and inputs.

How do you write a prompt that removes the defaults and checks the page?

Anthropic's instruction is short: name the patterns, then 'check which styles the first result used instead, and extend the list if needed'. A ban alone only moves the default, so each ban below is paired with a replacement. The review uses 1440 and 390 pixels, the widths Wooldridge checked, and adds tests for failures seen in these write-ups. The full Playwright setup and critique rubric are at /blog/visual-feedback-loop-claude-screenshots-playwright-ui-qa, and art-direction templates at /blog/claude-ui-design-prompts-art-direction-templates.

What goes wrong on Claude-built landing pages?

The praise is real. Wooldridge calls Fable 5.2 'the best Claude model I have used for UI design so far', and Muzli's verdict on Opus 5.5 was 'the first model I would hand a real brief'. Velox's counterweight: every page was a good first draft in under 12 minutes, and none was finished. These failures were reported in the first two days.

What should you check before a Claude-built page goes live?

The model's review does not replace yours. Baukys' Opus 5.5 run listed what it could not check, including how the scroll feels on a real phone.

What do we still not know about Fable 5.2 for design?

Treat this comparison as provisional and rerun it on your own brief.

• •No blind test exists where people rate pages without knowing the model. Every quality call above is the tester's own.
• •Velox tested Medium effort only and has a commercial interest. The which-ai data is one run per condition at xhigh. Wooldridge is one person on launch day. These cannot be averaged.
• •Arena, an independent board, lists Fable 5.2 as '#1 in WebDev' at xHigh under new releases with no score shown, and Velox reported on 10 November that this was an early estimate with no votes yet. We could not read Design Arena's ranking, so we do not cite it.
• •Anthropic has not said which model runs Claude Design, so this applies to Claude Code and API work. For that product's changes, see /blog/claude-design-standalone-closing-december-14-migrate-claude-motion.

Where Tech Arion fits

Tech Arion builds websites with Claude Code and a person reviewing the design. Our landing page work starts with the brand guide and content file described here. Details are at /services/landing-pages.

Frequently asked questions about Fable 5.2 and Opus 5.5 for web design