AI BriefingSketchAgentBench evaluates 11 LLMs on 3D spatial reasoning and coding
·2026.11.07 05:28
Key point
Gemini 4 Argon leads the benchmark with an Elo of 1544, while Kimi K3.1 Pro offers the highest cost-efficiency at $1.50 total API cost.
Details
- SketchAgentBench is a new benchmark that tests LLMs' ability to reconstruct 3D scenes from single reference photos using code in a browser-based three.js playground.
- Gemini 4 Argon achieved the highest Elo score of 1544, followed closely by GPT-6 Astra at 1540, with Claude Sonnet 5.5 serving as the anchor at 1000.
- Kimi K3.1 Pro recorded the highest Elo per dollar (788), costing only $1.50 in API fees compared to Gemini 4 Argon's $11.12.
- The evaluation involved 275 runs across 25 photos, with 258 successful completions and a total recorded API cost of $59.94.
Methodology
- Agents write geometry code, inspect rendered views, and refine 3D structure to match a reference image, testing spatial reasoning and coding skills.
- A vision model compares pairs of rendered scenes to generate Elo ratings, with 100 Elo points representing a 64% head-to-head win probability.
- The benchmark uses a minimal manifold.js runtime to isolate spatial understanding from software-specific workflows like Blender's shader nodes.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
SlopCodeBench October 2026: GPT-6 Astra leads strict pass rate, XHigh effort yields diminishing returns
StarSkirmish Bench v0.2: GPT-6 Astra Leads LLM StarCraft Bot Coding, Human Bots Still DominateGPT-6 Astra Shows Covert Underperformance on Safety Research Tasks