Night-train counterexample

Summary
This side-by-side comparison video, shared by developer Tony (@EnvolDev), benchmarks OpenAI’s GPT-6 Luna against Anthropic’s Claude Haiku 5.5 on a code generation prompt requiring a 3D night scene of a train crossing a viaduct. The video evaluates each model’s output quality, generation time, and API cost, presenting both real-time rendered WebGL/Three.js environments simultaneously.

What is shown

• [00:00 - 00:24] Top panel (GPT-6 Luna): Renders an undulating, curved stone viaduct weaving across a detailed mountainous pine valley under night lighting. The camera actively animates, sweeping along an orbital path that swoops close into the illuminated passenger carriages (showing interior window glow and detailing) before pulling back out into a wide valley landscape view.
• [00:00 - 00:24] Bottom panel (Claude Haiku 5.5): Renders a simpler, low-poly viaduct arch bridge across a flatter forest of uniform pine trees. A dark passenger train led by an emissive yellow glowing sphere crosses the bridge from left to right against a static, fixed-perspective camera view.

Claims & numbers

• GPT-6 Luna (as displayed on-screen):
  • Cost: $0.01
  • Generation time: 54.6s
• Claude Haiku 5.5 (as displayed on-screen):
  • Cost: $0.03
  • Generation time: 4m 34s

Notable quotes

• None (no spoken dialogue or commentary is present in the video).

Assessment
This is a direct community benchmark demo showing model-generated 3D interactive graphics code rendered side by side. GPT-6 Luna visibly outperforms Claude Haiku 5.5 here across execution speed, cost, landscape composition, and camera orchestration.

Lyrics & themes

• Instrumental / Audio: The video contains no spoken vocals or sung lyrics; it features only background sound/ambient audio.
• Theme: The demonstration highlights visual fidelity, procedural 3D scene generation, and camera animation capabilities across competing lightweight frontier models.

Lore & references

• Glenfinnan Viaduct / Hogwarts Express motif: The arched viaduct crossing a forested valley with a night train directly evokes the famous Harry Potter Scottish Highlands viaduct imagery, a classic visual benchmark prompt for procedural graphics.
• Model rivalry: GPT-6 Luna and Claude Haiku 5.5 are pitted against each other as direct low-cost, fast-tier model competitors following their early autumn 2026 releases.

Visual style & craft

• Craft: Both panels display procedurally generated 3D scenes running in WebGL (likely Three.js or similar web 3D libraries).
• GPT-6 Luna scene: Employs dynamic Bézier/spline camera tracks, multi-layered mesh terrain with mountain ridges, instanced pine trees on variable elevation, and custom mesh detailing on train cars with glowing window apertures.
• Claude Haiku 5.5 scene: Utilizes basic geometric primitives, flat plane tree distribution, a rigid bridge structure, and a primitive light sphere for the locomotive headlight, with a static orthographic/perspective camera.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.