AI coding agents can now watch a real video and attempt to reconstruct the entire scene in Blender using Python. But which model actually understands the 3D world behind the pixels? Meet BVB, the Blender-VideoBench, a new open benchmark that tests AI agents by making them rebuild real-world videos as animated 3D scenes. No photogrammetry shortcuts. No downloaded 3D assets. No ready-made furniture models. Just video frames, Blender primitives, Python code, and a fixed budget. HOW DOES BVB WORK? • 288 real indoor videos: scenes from ARKitScenes, ScanNet, and ScanNet++. • 5,130 spatiotemporal questions: testing objects, distances, directions, room geometry, and camera movement. • Identical conditions: every agent uses the same Mini-BVB harness, sandbox, instructions, and cost limit. • Programmatic reconstruction: models build actual Blender scenes using geometry, materials, objects, and animated cameras. • No external asset libraries: every object must be constructed using Blender's own primitives and operators. • Dual VQA: measures how well reconstructed videos preserve facts from the original footage. • Latent Similarity: compares spatial layout and movement using a frozen video representation model. • Failed reconstructions count: unsuccessful scenes receive zero scores rather than disappearing from the results. CURRENT BVB LEADERBOARD Top configurations by overall score: 1. Claude Fable 5.2 — 76.44 2. GPT-6 Astra — 76.33 3. GPT-6.1 Sol — 75.41 4. GPT-6.1 Terra — 75.16 5. GPT-6 Astra (high) — 73.95 Gemini 4 Argon and Claude Opus 5.5 are also among the leading configurations. And here's an interesting surprise. Inkling 2 Small, from Thinking Machines Lab, reaches 71.56 points at an average agent cost of just $0.257 per scene. Meanwhile, the top-ranked Fable 5.2 configuration costs approximately $2.61 per scene. So model size and raw reasoning power aren't the only things worth comparing. Cost efficiency matters too. THE BIGGER QUESTION Does reconstructing a video mean the AI actually understands the scene? A model might create a visually convincing room while getting the object distances, proportions, or spatial relationships wrong. That's why BVB evaluates both visual similarity and factual spatial consistency. And that's what makes this benchmark different from another image-generation leaderboard. We're testing the ability of AI agents to infer a 3D world, construct it programmatically, and render it from a moving camera. This could become increasingly important for automated 3D production, virtual environments, game development, and AI-assisted VFX. But these scores don't mean the resulting scenes are production-ready or geometrically accurate. BVB is a controlled research benchmark, not a complete test of Blender modeling, art direction, or professional 3D workflows. The original paper evaluated 51 configurations. The live leaderboard has since expanded and continues to change as new models are added. Explore the interactive benchmark and reconstructions: https://yoloytang.me/BVB/ GitHub, code and evaluation tools: https://github.com/yoloytang/BVB Research paper: https://arxiv.org/abs/2609.15478 Which model would you trust to reconstruct a real-world scene in Blender? #BlenderAI #ClaudeFable52 #AIBenchmark