One Wish, One Racetrack: DeepSeek's Wish-Driven Programming Builds a Racing Game for 3 RMB

Content written on 2026.10.25

DeepSeek-V4.1-Flash-1017 has been out for a while now. Although it's been used as one of the three flagship models for a long time, I never had a chance to study it seriously.

In particular, the "Wish-Driven Programming" proposed by DeepSeek and several of its official examples are quite impressive.

What exactly did DeepSeek-V4.1-Flash-1017 upgrade?

First, look at the official data.

On October 17, DeepSeek released the V4.1-Flash-1017 update.

Several metrics worth noting:

Although model benchmark scores today are just for reference, which indicators have significantly improved still shows where model capabilities have changed.

The improvement in Terminal-Bench and Automation-Bench suggests it's worth paying attention to terminal operations, tool execution, and automation tasks. The improvement in ProgramBench reflects progress in programming problem-solving ability.

These capabilities are all related to the Harness. Or rather, this update was itself a targeted optimization for DeepSeek Harness, so that DSH could perform better.

Of course, given the close relationship between Tencent and DeepSeek, WorkBuddy and CodeBuddy are also on the optimization list.

More noteworthy than benchmark scores: performance–cost Pareto improvement

The official side emphasized a new term this time: performance–cost Pareto improvement.

Simply put: although the token unit price for DeepSeek-V4.1-Flash-1017 hasn't changed directly, the number of tokens required for a single task is being optimized and reduced as much as possible, thereby achieving a real reduction in actual needs.

For agents, this is actually quite important.

As agents can complete longer and longer tasks, each agent task is not a single request with a single response. It's a process involving numerous model requests, numerous tool calls, and numerous loops. Because this process is so complex, everyone has been unconsciously skipping it.

This time, DeepSeek is targeting it.

DeepSeek hopes that by optimizing the agent process, reducing invalid output and shortening the time needed to complete tasks, it can achieve lower actual usage costs.

From this perspective, Liang Sheng, still our Liang Sheng.

Replicating the official racing game with DSH + official API

I've always liked racing games, especially Need for Speed, so let's try to replicate the official racing game.

The environment is the specifically optimized DeepSeek Harness (DSH), calling DeepSeek-V4.1-Flash-1017 through the official API.

First attempt: directly make a wish

For the first attempt, I really used "wish-driven" programming.

I didn't prepare a detailed requirements document, nor did I design a technical solution in advance. I directly gave one sentence:

Make me a racing game, that can drive, has a track, has timing, and supports both a driver's view and a rear-car view.

If you just look at the homepage, it looks pretty good.

After entering the game, the basic gameplay was already implemented: you can drive, there's a track and timing, and it supports switching between driver's view and rear view.

But the overall visual style and level of detail of the game were some distance from my expectations for a racing game.

Second attempt: adjusting the prompt

Although the first result wasn't that satisfying, I felt the model still had potential.

So I asked the AI to compare the official example with my result and give suggestions for improving the prompt.

This time the effect was clearly much better.

The prompt was as follows:

Please use Three.js or native WebGL to develop a single-file HTML realistic night-rain city style 3D racing game, referencing the night city track from Need for Speed. Specific requirements: Scene: late-night city road, wet reflective road surface, tall buildings and streetlights on both sides. The track has red-and-white curbs and metal guardrails on both sides. Vehicle: rear view, generate a detailed sports car (with spoiler, tail lights, license plate). Support arrow keys for left/right steering, up/down for acceleration and braking, with steering inertia and a slight rear squat effect when accelerating. HUD: upper left shows current position (e.g., P4/4) and lap (e.g., 02/03); top center shows real-time lap time and best lap time; lower right shows speedometer (number + arc pointer); lower left shows track minimap, overall like a professional racing game UI. Gameplay: simple collision detection (slowing down when hitting guardrails), after crossing the finish line display total time and position.

The overall visual style and completeness of the game improved significantly, coming much closer to what I wanted.

Runtime and token consumption

The first attempt was too simple, so we won't pay attention to it.

The second command took 51 minutes 14 seconds, 60.1M tokens, and cost about 3 RMB (with two tasks running simultaneously, the other being slightly smaller).

I can only say deep seeking really saves money~