Frontier3DBench: coding agents ace single 3D gizmos (83) but fail full worlds (30)
机器之心 · wechat · 2026-11-06
Einsia AI released Frontier3DBench, the first 3D-creation benchmark for coding agents spanning both Blender and Unreal Engine: 9 categories, 145 tasks (65 objects, 65 scenes, 15 open worlds), scoring 11 models on appearance, structure, and behavior.
Key findings: GPT-6.1-sol tops the overall leaderboard at 66.02, but GPT-6-astra leads open worlds (61.28), Claude Opus 5.5 leads text-to-object (84.05), and MiniMax-M3.1-Flash is the scene-editing dark horse (78.70). Every model scores far lower on reference-image reconstruction than free generation (76-84 vs. max 69.61 on multi-view). The biggest gap is in open worlds: local interactions average 83.08, but motion obstacle avoidance drops to 20.82 and state/workflow to 30.26—agents can build single gizmos but can't chain them into working worlds. On cost, GPT-6-luna ($0.27/task), DeepSeek-v4.1-Flash ($0.45), and Sol ($1.01) form the Pareto frontier; Opus costs $21.17/task yet scores below Sol.
Related event: Frontier3DBench Tests Coding Agents on 145 3D World-Building Tasks(2 posts)→
More from coding & agent
- 7 new MCP servers this week, including one that phones you when your agent gets stuck — Prior_Engineering486 · 2026-11-07
- Dev replaces vector-search RAG with Markdown + knowledge graphs via Claude CLI, claims ~95% fewer hallucinations — Shivam__kumar · 2026-11-07
- RecallDesk: a support agent with persistent memory that separates attempted fixes from confirmed resolutions — lakshmisrivalli · 2026-11-07
- When two agents agree the job is done but the customer still gets the wrong product — krishnan · 2026-11-07
- XState v6 beta lands: leaner API, stronger types, durable execution, built for AI agents — DavidKPiano · 2026-11-07
- MCP tracing + send_feedback instantly revealed how agents really use an MCP server — indragie · 2026-11-07