# Kimi vs. Claude vs. GPT-6 Astra: AI Design Performance Across Five Tasks The [Design Arena](https://designarena.ai) leaderboard measures AI model performance on visual and creative tasks using human preference voting: websites, 3D scenes, UI components, browser games, and data visualization. The rankings reveal something more interesting than a single winner. Different models lead on different tasks, and the gaps between them vary significantly by category. This article runs five creative challenges through Kimi K3, Claude (Fable 5.1), and GPT-6 Astra, comparing their outputs and drawing practical conclusions about which model to use for what. ## Round 1: Web design **The challenge:** Design a complete landing page for "Conveyor," a PostgreSQL-native job queue. Target audience is developers. Required elements: hero section, feature explanations, pricing table, calls to action. ### Claude Claude's design led with an animated hero graphic that showed items moving between states in a job queue (deferred → running → done), demonstrating genuine product understanding. Layout logic was sound and supporting graphics were competent. ![A dynamic visualization of a job queue, showing tasks moving between states like 'deferred', 'running', and 'done'.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/6d6dba2d-e327-47c0-2315-9c167e2d9600/md2x =1920x1080) The weakness: typography. The headline font felt slightly off for a modern developer tool, somewhere between dated and generic, which undermined the polished graphics around it. ### GPT-6 Astra Astra's typography was a clear improvement. Font choices were cleaner and more appropriate for the developer audience. Where it fell short: the graphical elements. The hero visual was static, and the supporting charts and diagrams were notably less refined than Claude's. ![A side-by-side comparison of charts, with Claude's version (top) appearing more polished and professional than OpenAI's (bottom).](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/79081f23-5dcc-426d-9e49-13ab406faf00/orig =1920x1080) ### Kimi K3 Kimi produced a production-ready page. Typography and hierarchy were well-balanced, the animated hero felt integrated rather than bolted on, and UX details were present: the `npm install conveyor` code snippet had a one-click copy button, a small but telling signal that the model understood its audience. ![The hero section of Kimi's winning landing page design for "Conveyor," showcasing excellent typography and a clean layout.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/15270e46-56ce-4e4f-0636-48bccd012500/lg2x =1920x1080) **Winner:** Kimi K3. This result is consistent with Design Arena's current leaderboard: Kimi K3 holds the top spot at 1361 Elo on the Website category, ahead of Muse Spark 1.3 (1343) and Claude Fable 5.1 (1330). **Practical note:** For developer-focused landing pages, specify your target audience explicitly in the prompt, request interactive or animated elements to visualize the product's core function, and define the aesthetic (clean, dark mode, information-dense). Kimi consistently delivers on these constraints. ## Round 2: 3D modeling **The challenge:** Generate an animated, low-poly floating island with a day-to-night lighting cycle using Three.js. GPT-6 Astra won this round by a meaningful margin. Its composition felt more cohesive: the island geometry was less blocky, the floating elements were better distributed, and the day-to-night lighting transition was gradual and naturalistic rather than harsh. Kimi produced the weakest result of the three, with geometry that felt rough and a lighting transition that lacked the subtlety of Astra's version. Claude landed in between. **Winner:** GPT-6 Astra. The current Design Arena 3D leaderboard confirms this: Astra leads at 1481 Elo, with Kimi K3 second at 1426 and Claude Fable 5.1 third at 1423. Astra's 55-point lead translates to winning roughly 58 out of 100 head-to-head votes against Kimi. Astra's 3D advantage is consistent with its BenchCAD score of 95.9%, substantially ahead of GPT-5.6 Sol (83.3%) and Claude Fable 5.1 (84.3%). Spatial reasoning, material handling, and lighting coherence appear to be genuine Astra strengths. ## Round 3: UI component library **The challenge:** Build a complete, cohesive design system with multiple components. All three models performed strongly here. Each produced a comprehensive set of components, all technically complete and visually consistent. The differences were aesthetic: Claude's components leaned functional and clean; Kimi's felt slightly more polished with stronger visual hierarchy; Astra's fell between the two. **Winner:** Three-way tie. The right choice depends on which aesthetic matches your existing design system. ## Round 4: Mobile app design **The challenge:** Design five screens for a coffee rating and tracking app: dashboard, brew detail, add brew form, bean list, settings. All three models failed this challenge. The outputs were universally generic: blocky layouts, default-looking components, no design personality. None of the results were close to shippable. **Winner:** None. ![A side-by-side comparison of the coffee rating apps from all three models, highlighting their generic and unpolished appearance.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/d1aa9cf0-318b-4aba-77db-438b41ccac00/lg2x =1920x1080) The failure pattern is consistent with what makes multi-screen mobile design hard: it requires coherent user journey mapping, information hierarchy across different contexts, and constraint-aware layout thinking, all at once from a single prompt. The current approach that works better: prompt for one screen at a time, generate individual components separately, and assemble them with human oversight. Don't ask any of these models to design a full mobile app in one shot. ## Round 5: 2D platformer game **The challenge:** Generate a playable 2D side-scrolling platformer with a robot character, walking and jump animations, and a hover/jetpack ability. ### Claude Claude produced the strongest result. The character design was detailed and coherent, with a subtle blinking eye animation that added personality. Walking animations were fluid, the jetpack flame effect was well-executed, and the background created an atmospheric setting that complemented the character art. The game felt like a thoughtful prototype. ![The playable 2D platformer game generated by Claude, showing the detailed robot character in a moody, atmospheric environment.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/1c8b3f8f-ce35-4107-f018-826a485bb000/lg1x =1920x1080) ### GPT-6 Astra Astra's entry had significant technical problems. The walking animation appeared to run backward, the jetpack effect was minimal, and the overall feel was janky. A clear step down from Claude. ### Kimi K3 Kimi's game landed in the middle. Animations worked correctly, and the robot's eye tracked the user's mouse cursor, a nice interactive touch. Overall art style and animation quality were less refined than Claude's. **Winner:** Claude. Its ability to generate a coherent set of assets, animations, and atmosphere demonstrates real capability for rapid game prototyping. For indie developers testing a game concept quickly, Claude is currently the most reliable option among these three. ## Summary | Task | Winner | Notes | | :------------ | :------------ | :-------------------------------------------------------- | | Web design | Kimi K3 | Production-ready; leads Design Arena Website at 1361 Elo | | 3D modeling | GPT-6 Astra | Leads Design Arena 3D at 1481 Elo; BenchCAD 95.9% | | UI components | Three-way tie | Aesthetic preference drives the choice | | Mobile app | None | All models failed; prompt individual screens instead | | 2D platformer | Claude | Coherent assets and animations; weakest result from Astra | The main takeaway from these tests is simple: **there is no single best model across all five design tasks**. Each one has a clear strength, which makes a multi-model workflow more practical than trying to force one model into every use case. **Kimi K3 is the strongest choice for web design and landing pages**, while **GPT-6 Astra stands out for 3D work**. Claude performed best on interactive experiences and game assets, and for UI components, the choice is mostly aesthetic since all three produced competitive results. The mobile app test was the outlier. None of the models handled the full app well in one pass, so **prompting individual screens separately is the better approach** for that kind of work. Cost is also becoming less of a barrier. Running all five Kimi demos in these tests came to roughly **$6 in total**, which shows how affordable exploratory design generation has become, even at the premium end of the model market.