Running Qwen 3.8 27B (Q4, llama.cpp) on a single RTX 3090 required about 114 minutes to code three small 3D games solo. When the same model acted as a worker under a cloud‑based GPT‑6.1 Sol orchestrator that split and verified the workload, the total time dropped to roughly 43 minutes, while the orchestrator alone needed ~18.6 minutes.