A developer tested ten different model and harness combinations to generate a Three.js sci-fi hangar scene featuring hovering drones, warning lights, and animated runway strips. Using /goal mode with an identical prompt across all setups, results varied significantly in generation time and output quality. The fastest completion came from GLM 5.3 Flash Max Codex Open9m, finishing in 0.232 seconds with 143,852 output tokens and a 14.38% block rate.

In contrast, Qwen 3.8 27B x-high harnesses like Open Code and DSH required 8 to 41 minutes, producing between 665 and 3.4 million tokens. Results showed a trade-off between speed and token volume, with GLM Flash models prioritizing efficiency and Qwen models maximizing output length. All GLM runs used the Flash Max Codex configuration, while Qwen experiments utilized x-high settings and different execution environments.

The test provides a comparative benchmark for developers evaluating AI coding assistants on complex Three.js prompts.