DeepSeek vs Gemini Argon vs GPT-6 Astra: the same racing game prompt, and DeepSeek’s run cost $0.58

I gave DeepSeek V4.1 Flash, Gemini 4 Argon and GPT-6 Astra the same reference image and prompt: build a playable low-poly racing diorama with automatic laps, keyboard controls and engine sound. Each got a separate project and one initial build prompt. Their CLI harnesses handled the implementation, screenshot reviews and revisions. No game code was hand-edited between runs. DeepSeek’s result makes a pretty good case for its price. It produced a complete driving scene with an elevated crossing, animated wheels and synthesized audio for roughly 58 cents in estimated API usage. All three games passed external checks for three automatic laps, keyboard driving and recovery. Model Harness Effort Time Tokens* Cost DeepSeek V4.1 Flash AgenC Core 0.19.27 Max 53m 38s 41.71M ~ $0.58 API Gemini 4 Argon Gemini CLI 0.29.3 High (maximum) 22m 47s 11.70M ~$2.58+ API GPT-6 Astra Codex CLI 0.164.0 Max 45m 02s 2.38M ~$6.54 API Disclosure: we work on AgenC Core . DeepSeek used our harness; the other two used their native CLIs. This compares the complete setups on one task, so the harness matters alongside the model.