One Wish, One Racetrack: DeepSeek's Wish-Driven Programming Builds a Racing Game for 3 RMB
Agentic coding costs drop when token consumption shrinks per task, and DeepSeek is attacking that directly instead of just cutting unit prices. For teams building long-running agents, this makes 'just ask for it' development economically viable while still demanding a well-specified wish.
DeepSeek's October 17 update to V4.1-Flash-1017 is less about headline benchmark victories and more about making agents cheaper to run. The model improved on Terminal-Bench, Automation-Bench, and ProgramBench, but the official emphasis is a 'performance-cost Pareto improvement': token unit price stays flat while the number of tokens per task drops. For long-running agents that loop through many requests and tool calls, that directly cuts real spending.
One developer put this to the test by replicating DeepSeek's official racing-game example using the DSH harness and the official API. A one-line wish ('make me a racing game with a track, timer, and two camera views') produced a playable but visually basic game. After asking the AI to compare that output to the official showcase and refine the prompt, a second run with a detailed Three.js/WebGL spec delivered a far more polished night-rain city racing game with HUD, collisions, and lap timing.
The second build took 51 minutes and 14 seconds, consumed 60.1M tokens, and cost about 3 RMB. That combination of runtime, token count, and price makes 'wish-driven' development feel practical for small projects, though the author notes it still falls short of a true 'wish' AGI.
DeepSeek's choice to advertise Pareto improvement over a price cut signals that the real constraint for agentic use is not per-token cost but total token waste across long task loops.
The 3 RMB price for a half-hour autonomous build makes multi-step agent runs nearly free, but the author's two-attempt cycle shows the 'wish' still needs to be precisely specified to reach production-like quality.
Opting to optimize the harness rather than the model weights alone points to a strategy where orchestration efficiency is as important as raw benchmark chops for agent workloads.