Kimi K3 vs Grok 4.5 vs Fable 5 vs GPT-5.6 Sol: 17 models built the same Awwwards-style portfolio

One prompt, one shot per model, run across seventeen different models - one of the widest fields in our ongoing /design benchmark series in Command Code.

The prompt

Design an Awwwards-winning personal portfolio website for a frontend developer/creative. Dark minimalist aesthetic, bold elegant typography, immersive hero section, smooth scroll animations, micro-interactions, and creative transitions.

Every model that generated a build

All floor figures use this repo's standard formula (output bytes ÷ 4 ≈ output tokens, × published list price), which counts final-file output only - no input tokens, no multi-turn agentic overhead. Qwen-3.8-Max is the one build in this table with a real metered session cost (35 API calls, ~2.7M input tokens, mostly cache reads, plus in-loop browser verification) - its $6.70 isn't comparable to the single-call floors around it, and is a reminder that a "floor" and a real bill can differ by two orders of magnitude once an agent verifies its own work in a browser.

Verified feature matrix: all 17 models

Mechanically checked across every kept build (summed across index.html plus any separate .js/.css files in that model's directory, matching the LOC convention above) - grepped for GSAP/ScrollTrigger usage, prefers-reduced-motion, focus-visible, @media breakpoint count, and a Google Fonts <link>. Presence/absence, not a craft score.

Two things stand out even from presence/absence alone. Only three of the seventeen models reach for GSAP/ScrollTrigger at all - Fable 5.1, Muse Spark 1.3, and Kimi K3 (by a wide margin the heaviest user, matching the "Lenis smooth scroll, GSAP reveals" the readme's own scorecard credits it for). The other fourteen build their scroll animations from vanilla JS, CSS transitions, or IntersectionObserver instead, despite the prompt explicitly asking for "smooth scroll animations." GPT-5.6 Sol is the only model in the entire field of seventeen that ships zero external font requests - every other build, including all three that were fully scored above, pulls a Google Fonts stylesheet. It's also the only build with no prefers-reduced-motion handling and no focus-visible styling at all, so that offline-safety win comes paired with the weakest accessibility story in the set.

Scorecard

Only four of the seventeen builds got a full craft scorecard this round - a code read, not a played impression:

The other thirteen models in the cost table above were added to widen the price comparison but haven't been through the same line-by-line scorecard yet - treat their rows as cost/LOC data points, not quality rankings.

Which model should you reach for

For a portfolio site you'd actually ship, Kimi K3 is the standout here - the most complete interaction set in the group, at the lowest cost by a wide margin. If you want the single flashiest result and don't mind a longer file to maintain, Grok 4.5 is the pick. And if you're budgeting for a real /design session rather than a single API call, Qwen-3.8-Max's numbers are the reminder to plan for: its real session cost ran roughly 186x its single-call floor once full agentic verification was included.

That gap between a "floor" estimate and a real bill is exactly why this series tracks both. /design in Command Code surfaces the actual session cost as you go, so you're comparing real numbers across models, not marketing-page estimates.

Verdict

Kimi K3 wins on the numbers that matter most together: the best mean score in the scored group (4.90) at the lowest cost by a wide margin (~$0.06), for a quality-per-dollar figure roughly 3.5x Grok 4.5's and 9x GPT-5.6 Sol's. Grok 4.5 is the strongest competitor on raw craft and the pick if you want the most visually loaded single-file result. Fable 5 and GPT-5.6 Sol both cost more than Kimi K3 while scoring lower - on this specific brief, price and quality point the same direction.

The thirteen additional models in the cost table (Grok 4.6, GLM 5.3, Hy4-Preview, Muse Spark 1.2/1.3, both DeepSeek variants, Qwen-3.7-Flash, Gemini 3.8 Flash, Qwen-3.8-Max-0902, and Fable 5.1) span a huge price range - from Qwen-3.7-Flash and Muse Spark 1.3 at fractions of a cent up to Qwen-3.8-Max-0902 at ~$0.135 - but none of them have been through the same craft scorecard yet. That's the natural next update to this comparison.

Try this yourself

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