3DSceneBench: coding agents rebuild livable 3D rooms but top functionality score is just 42.1
yizhu59 · x · 2026-11-14
PranaLabs introduces 3DSceneBench, a benchmark asking whether AI-rebuilt 3D worlds actually work, not just look right: can you open the drawer, carry the cup, run the centrifuge?
Design
- Coding agents rebuild whole rooms from photos; hidden tests check if the single submitted world supports every task the room should enable
- 49 whole-room scenes across 7 categories, 14,219 objects, 471 tasks
- All reference worlds are simulation-ready and interactive, framed as a first step toward a spatial-intelligence playground
Results (agents on 10 frontier models)
- Best Appearance: 79.3/100 (GPT-6 Astra); best Geometry: 62.1 (Gemini 4.0 Argon)
- Best Functionality: only 42.1 (Claude Opus 5.5); leading agents fully complete roughly a quarter of tasks
Takeaway: frontier models can make worlds that look right but far fewer that work. More challenging scenes are teased for December.
Related event: 3DSceneBench Tests Whether Coding Agents Can Build Functional 3D Worlds(2 posts)→
More from coding & agent
- Superconductor adds Loops MCP connector so coding agents can manage email campaigns — sergeykarayev · 2026-11-14
- Dev records 2 hours of manual Neovim coding to get PRs accepted — yacineMTB · 2026-11-14
- Grok Bot tests auto-approving purchases with $50-$5,000 spend limits — XFreeze · 2026-11-14
- Cloudflare: AI pentesting framework traffic surged 30x; a six-pillar "Modern Security Model" to fight back — Cloudflare Blog · 2026-11-14
- Musk endorses Grok Bot prompt that turns everything you read into a self-organizing second brain — elonmusk · 2026-11-14
- Scoble backs Boardy agent, says autonomous 'principals' are the next phase of AI agents — Scobleizer · 2026-11-14