With only one scored task and a 75% confidence read, this is a lean, not a blowout. But the result is still straightforward: Kimi K3 won the only thing these models were asked to do, and it won for concrete reasons visible in the output.
In the Three.js challenge, Kimi K3 produced a scene that actually reads as the assignment: a rain-soaked nighttime gas station with a visible store interior, pumps, a sedan, reflections, and a coherent forward-moving camera. It was hardly elegant — the geometry was blocky, the scene sparse, and the car became too dominant late in the framing — but it was complete, stable, and recognizably on brief.
Claude Opus 4.8, by contrast, failed on the basics of visual communication. Parts of the theme were present, but the scene was so underlit and poorly composed that sampled frames were near unreadable, including an effectively black final image. That is not a minor aesthetic miss; it breaks the cinematic brief and undercuts the technical claim of a "complete, runnable" result.
The score gap, 6.5 to 2.1, reflects that difference in usefulness. Kimi K3 gave you something you could inspect, iterate on, and plausibly ship to a next draft. Claude Opus 4.8 gave you a scene that too often hid its own work.
Final call: Kimi K3 wins, narrowly in statistical confidence but decisively on the evidence we actually have. When one model delivers a readable gas-station sequence and the other fades into near-black frames, the editorial verdict is simple: Kimi K3 was the only model that really completed the task.
How they were tested
We ran 1 fresh anim task, generated on the fly for this matchup so neither model could prepare in advance, and had gpt-5.4 score each one. To cancel position bias, every task was judged twice — once in each presentation order — and every number reported here, including the headline totals, is the average of both passes. Kimi K3 scored 6.5 to Anthropic: Claude Opus 4.8's 2.1.
1. Create a complete, runnable Three.js
Create a complete, runnable Three.js scene designed specifically for a silent 10-second evaluation recording. The scene must begin instantly, animate on its own, and require zero keyboard, mouse, touch, or camera input. Build a highly realistic rain-soaked roadside gas station at night, seen through one carefully directed cinematic camera shot. ### First 10 seconds The opening frame must already look finished and visually striking. No screen, fade-in, setup view, camera teleport, or empty geometry. Direct the camera automatically: * 0–3 seconds: Slow forward dolly toward the station from street level. Keep the dark-red sedan prominent in the foreground and the illuminated convenience store visible behind it. * 3–7 seconds: Gently arc around the sedan, revealing wet bodywork, detailed fuel pumps, the person under the canopy, and the store interior through the glass. * 7–10 seconds: Ease toward a close three-quarter composition as lightning briefly illuminates the clouds, station roof, rain, and puddles. End on a strong frame. * Use smooth spline-based motion with eased acceleration. Avoid sudden turns, clipping, excessive camera motion, or a generic orbit around the scene. * Automatically aim the camera at changing points of interest so the composition feels deliberately directed. * Continue with a subtle looping cinematic camera path after 10 seconds.
Winner: Kimi K3 — Model A clearly reads as a rain-soaked nighttime gas station with visible store interior, pumps, sedan, reflections, and a coherent forward-moving camera, though it remains quite blocky and the car dominates too much of the later framing. Model B matches parts of the theme but is severely underlit and poorly composed, with near-unreadable frames and an effectively empty final image that undermines both the cinematic brief and technical polish. (Second judge pass, order swapped — scores are the average of both: Model A clearly reads as a rain-soaked nighttime gas station with a sedan, pump, person, lit storefront, and coherent camera progression, while Model B is extremely dark, partially unreadable, and even includes a near-black frame that breaks the brief. Model A is still stylized and somewhat sparse rather than highly realistic, but it is far more complete, purposeful, and technically stable across the sampled frames.)
See every prompt and the full side-by-side outputs in the interactive Head-to-Head.