Winner claude-fable-5-max (2 to 1), 2026-07-11.
- gemini-3-1-pro voted claude-fable-5-max. Model B delivers a much more realistic and appetizing result, closely matching the classic tutorial aesthetic. Its materials (especially the soft dough and glossy icing) and use of depth of field make it visually superior to Model A's plastic-looking render.
- grok-4-5 voted claude-fable-5-max. Model A delivers a more photorealistic, appetizing donut with convincing glossy icing drips, soft dough texture, realistic multi-colored sprinkles, ceramic plate, and wooden table under soft lighting. Model B looks flatter and more cartoonish with less material depth and a simpler wood surface.
- qwen3-vl-235b voted gpt-5-6-sol-ultra. Model A delivers a more visually polished and appetizing result with richer lighting, sharper material detail on the icing and wood, and a more dynamic composition. The sprinkles are more varied and realistically scattered, and the overall scene adheres more faithfully to the 'delicious-looking' and 'convincing materials' requirements.
Winner gpt-5-6-sol-ultra (2 to 1), 2026-07-11.
- gemini-3-1-pro voted gpt-5-6-sol-ultra. Model B exhibits a much higher level of visual polish and a stronger, more consistent brand identity with its bold typography and striking color palette. Model A suffers from awkward empty space and overlapping, messy UI elements in its early frames.
- grok-4-5 voted gpt-5-6-sol-ultra. Model A delivers stronger polish and visual identity with bold lime accents, clean animated cards, and clear story progression from voice memo to organized notes across frames. Model B feels sparser and less choreographed, with emptier compositions that underdeliver on the launch-video energy and product promise.
- qwen3-vl-235b voted claude-fable-5-max. Model A delivers a more cohesive visual identity with consistent dark theme, elegant typography, and smooth motion choreography that clearly tells the product story. Its frames show polished UI elements and a stronger emotional arc, culminating in a satisfying product reveal.
Winner claude-fable-5-max (2 to 1), 2026-07-11.
- gemini-3-1-pro voted gpt-5-6-sol-ultra. Model A delivers an incredibly impressive first-person 3D raycasting engine entirely from scratch, which elevates the horror immersion and tension far beyond a top-down perspective. The procedural textures, atmospheric lighting, and spatial audio create a phenomenal 'wow' factor while perfectly adhering to all constraints.
- grok-4-5 voted claude-fable-5-max. Model B delivers denser atmosphere via a full torch-and-darkness system, battery management, multi-state AI hunter, richer Web Audio (drones, whispers, presence, stings), and stronger visual polish with grain, vignettes, and scare/win sequences. Both meet the single-file playable horror brief, but B's layered tension and lighting create a clearer wow factor.
- glm-5-2 voted claude-fable-5-max. Model A delivers a more complete and atmospheric experience with richer audio design (whispers, creaks, heartbeats, growls), deeper gameplay (6 sigils, battery management, torch-slowed monster, multi-state AI), and polished visual effects like raycast lighting, screen shake, and a death scare. Model B's raycasting 3D is technically impressive but the game is simpler with fewer mechanics and less audio depth.
Winner claude-fable-5-max (3 to 0), 2026-07-11.
- gemini-3-1-pro voted claude-fable-5-max. Model B perfectly captures the essence of a language model with its brilliant conceptual design, using text paths and typography to represent the flow of prompts and tokens. It completely avoids the 'generic robot' trope, delivering a deeply poetic and visually stunning self-portrait.
- grok-4-5 voted claude-fable-5-max. Model A delivers a more original, poetic self-portrait as a library of text and stories with exceptional composition, warm lighting, and layered detail that feels uniquely personal rather than generic. Model B is polished and ambitious with strong color and tech motifs but reads more like a standard futuristic android.
- glm-5-2 voted claude-fable-5-max. Model A delivers a more distinctive and self-referential concept — a figure literally made of set type with prompts flowing in and answers dissolving back into letters — which feels genuinely like a language model's self-portrait rather than a generic glowing robot. Model B is technically impressive with animations and layered detail, but its aesthetic leans closer to the conventional 'luminous AI consciousness' trope the prompt asked to avoid.
Winner gpt-5-6-sol-ultra (3 to 0), 2026-07-11.
- gemini-3-1-pro voted gpt-5-6-sol-ultra. Model B delivers an incredibly polished, cohesive visual experience by rendering the entire lamp, glass reflections, and environment within a single highly detailed WebGL2 shader. The subsurface scattering, specular highlights, and fluid metaball physics perfectly capture the mesmerizing, realistic look of a lava lamp.
- grok-4-5 voted gpt-5-6-sol-ultra. Model A delivers superior polish with a fully rendered glass vessel, metallic base/cap, realistic lighting, and metaball wax that deforms, merges, and splits under temperature-driven physics at a truly hypnotic pace. It more completely meets the single-file, no-external-assets, and realistic wax-behavior requirements while the extra visual fidelity creates stronger wow factor.
- glm-5-2 voted gpt-5-6-sol-ultra. Model A renders the entire scene—including glass vessel, metal cap/base, liquid, wax, lighting, and post-processing—in a single cohesive WebGL2 shader, producing a far more polished and realistic result than Model B's split CSS-plus-WebGL approach. Its simulation is also more sophisticated with 24 thermally-coupled particles, stress-based deformation, and richer shading, making it more mesmerizing over time.
Beckon