Kimi K3 & Qwen-3.8-Max vs Fable 5.1 vs Opus 5: how the cheaper models out-built Claude on a scroll-driven 3D page

Four models, one prompt-thread, a real WebGL/Three.js scroll experience instead of a game this time. Part of our ongoing /design benchmark series in Command Code.

The prompt

Create the most creative experimental single-file HTML landing page for a multidisciplinary artist "Kael Thorn". Tailwind + Three.js. Full-page scroll journey where each section is a different "dimension": 2D → 3D transition on scroll, hero with massive 3D text that breaks apart into particles on scroll, an interactive timeline where scrolling warps space, portfolio pieces as floating 3D cubes you can spin, a philosophy section with text that assembles from fragments, and a final portal-like contact section. Push creativity to the maximum. Surreal, immersive, mind-bending design. Dark cosmic theme. Output the full HTML file.

Models and cost

Kimi K3 sent 3 prompts in the same thread (build → continue → "scroll feels jumpy, make it smooth"); Opus 5 and Qwen-3.8-Max were each a single /design invocation. Fable 5.1's figure is a one-shot floor from output bytes at Claude's $50/MTok output rate - the same rate this repo has used for Fable 5.1 elsewhere (chrome-dino, story-telling).

On the numbers that are actually comparable, Fable 5.1 is the most expensive build in the set despite shipping the least code - 602 LOC versus Opus 5's 1,277 and Qwen-3.8-Max's 1,205. Claude's $50/MTok output rate is 2x Opus 5's and roughly 17x Qwen-3.8-Max's implied rate, so a smaller file doesn't translate into a cheaper one.

Seen on X

A separate, shorter-prompt Kimi K3 vs Opus 5 head-to-head on the same general idea - a scroll-driven Three.js site with parallax and scroll-triggered animation - was posted with the actual clip:

NR

Naymur Rahman

@naymur_dev

Kimi K3 leads in scroll-driven Three.js performance What I tested - A scroll-based Three.js site - Compared Kimi K3 vs Claude Opus 5 - Built with Command Code /design feature (Kimi K3)...

3:34 PM·Jul 30, 2026

That's a different, shorter prompt than the "Kael Thorn" brief this post is built around, so treat it as a second, independent data point in the same direction rather than the same benchmark - but it's the same conclusion: Kimi K3 holding up better than Opus 5 once real scroll interaction is involved.

What the code actually shows

Fable 5.1 builds a single persistent Three.js scene - an earth and moon that travel along a hand-tuned scroll-progress path, with GSAP-driven parallax layers and text reveals for the four narrative chapters. It's clean, coherent, and respects prefers-reduced-motion throughout. But it's also the most restrained of the four on raw interaction: a grep for pointer/mouse event handlers and custom ShaderMaterial instances turns up 3 matches in the whole file.

Qwen-3.8-Max goes considerably further on the same brief. It builds six distinct Three.js scenes, one per "dimension" - a particle field that assembles into the artist's name from sampled canvas-glyph data, a wireframe terrain grid that rises out of flat 2D, a tunnel of rings and dust streaming past for the timeline, a draggable cube gallery driven by real raycasting (THREE.Raycaster, pointer-down/move/up handlers that let you grab and spin a portfolio piece), a converging particle sphere for the philosophy section, and a swirling portal vortex for contact - each with its own hand-written GLSL vertex/fragment shader pair. It also runs a custom hidden cursor with a magnetic-hover ring, and an on-screen HUD (dimension name, scroll percentage, a fake telemetry readout) that updates continuously.

Kimi K3 independently converged on the same core interaction - a raycasting-driven draggable cube gallery with near-identical math (dragging.rotation.y += dx * 0.008 in both builds) - and a grep for the same feature set (ShaderMaterial, ScrollTrigger.create, pointer/mouse handlers) turns up 8 matches in its file versus Fable 5.1's 3.

Verdict

On this prompt, the two cheaper builds are also the two with more ambitious execution. Qwen-3.8-Max, at a ~$0.065 floor, ships six custom-shader 3D scenes and a raycasting drag interaction that Fable 5.1 - at a $0.41 floor, the most expensive build in the set - doesn't attempt. Kimi K3 lands in between on cost ($0.24 for the full three-prompt session) and independently builds the same cube-dragging mechanic as Qwen-3.8-Max, down to similar rotation-scaling constants.

Fable 5.1's build is not bad - it's coherent, accessible, and the leanest file in the set - but on a prompt that explicitly asks to "push creativity to the maximum" with cubes "you can spin," it's the one build that doesn't ship the spin.

Opus 5 sits at the top of the cost table (~$0.37) for the most code (1,277 LOC) in the set; we haven't done a full line-by-line read of its interaction layer for this update, so treat its ranking here as cost-and-LOC only, not a quality judgment - see Limitations.

Limitations

• n=1 per model on this run (Kimi K3's 3-prompt thread was for smoothing scroll feel, not a re-roll).
• Feature counts for "ShaderMaterial / ScrollTrigger / pointer handlers" are mechanical grep matches, not a full line-by-line audit of every model - they're a reliable signal for interaction density, not a complete quality score. Opus 5's file was not read in the same depth as the other three for this post.
• Cost figures mix bases: Fable 5.1 and Opus 5 are one-shot floors from output bytes; Kimi K3's is a modeled full-session estimate; Qwen-3.8-Max's is a floor only and, per this repo's own metered runs elsewhere, likely understates real session cost once multi-turn agentic overhead is counted.

Try this yourself

Full command reference: commandcode.ai/docs/slash-commands/design. More one-shot comparisons: CommandCodeAI/slash-design-showcase.