Kimi K3.1 Pro vs Opus 5.5 vs GPT-6 Sol: A 3D Racing Game in 1 Hour

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Kimi K3.1 Pro vs Opus 5.5 vs GPT-6 Sol: A 3D Racing Game in 1 Hour · Channel: Zero to MVP · 05:45 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI models

Summary

The video compares Kimi 3.1 Pro, Claude Opus 5.5, and GPT-6 by asking each to build a playable 3D racing game in the same starter project within one hour. Kimi finished in 18 minutes and produced a cyberpunk-style game with a minor collision issue involving some obstacles. Opus took almost 40 minutes and, in the creator’s view, delivered the most polished-looking result. GPT-6 made a visually appealing game, but its steering controls were reversed because of a flipped sign that its tests did not catch. The creator concludes that Kimi performed well overall, while noting that the comparison is not conclusive.

Key points

The task was to build a third-person arcade racing game using a minimal project with 3.js, TypeScript, and Vite.
Each model had the same requirements and one hour to code, run, test, and fix its game; the creator did not intervene.
The game requirements included a closed-loop track, drivable car, obstacles, collisions, checkpoints, three laps, and a finish.
Kimi completed its game in 18 minutes and chose a cyberpunk visual style.
Kimi’s car could drive through obstacles that were slightly off to the side.
Claude Opus 5.5 took almost 40 minutes; the creator considered its game more polished, while noting that the visual comparison was subjective.
GPT-6’s steering was reversed because a positive heading angle in 3.js turns the car left; its tests did not cover controls.
The creator judged Kimi capable of competing in 3D game-building, but said it was too early for final conclusions.

Chapters

Start	Chapter	Summary
00:00	Kimi 3.1 Pro and the 3D game challenge	The creator introduces Kimi’s claimed capabilities and sets up a one-hour comparison to build a 3D racing game.
01:03	Shared requirements and test setup	The three models receive identical starter projects and requirements, including a track, car, collisions, checkpoints, and a finish.
02:06	Kimi’s result	Kimi finishes in 18 minutes with a cyberpunk-style game that works overall, although some obstacles do not reliably collide.
03:09	Opus and GPT-6 results	Opus produces a result the creator finds more polished, while GPT-6’s game has reversed steering caused by a sign error.
04:43	Comparison and conclusion	The creator says all three broadly completed the task, sees Kimi as a strong 3D competitor, and plans to evaluate it further.

Quotes

And all of it has to actually work, not just look nice in a screenshot. — 01:35

It finished in 18 minutes, which is pretty fast. — 02:06

To me, this version feels more like a finished game. — 03:41

So, the whole problem comes down to a single flipped sign. — 04:12

Visually, in my opinion, it’s weaker, but otherwise, it’s very good. — 05:14

Concepts and entities

Kimi 3.1 Pro, Claude Opus 5.5, GPT-6 Sol, Moonshot, Three.js, OpenCode, Unity, Unreal, Godot

Full transcript (original language)

00:00 Moonshot has just released Kimi 3.1 Pro, and according to the company’s own tests, it has almost caught up with Claude Opus 5.5. On Terminal Bench 4.0, the gap is less than a point, 65.7 versus 66.4. On top of that, a million tokens cost $2 for input and 10 for output. And Moonshot promises to release the open weights by November 4th. But what really caught my attention in the announcement was something else. Moonshot is making a

00:31 big bet on 3D. The model writes code for 3D models, works with Unity, Unreal, and Godot. And in internal testing, according to the company, it built several 3D games. How playable those games actually are, Moonshot doesn’t say. So I decided to find out for myself. For this, I prepared a dedicated task, build a 3D racing game, and it has to be done within 1 hour of work. I’m also giving the same task to Claude Opus 5. 5 and GPT-6 Soul to see how close

01:03 Kimi has come to the flagships. An hour from now, I’ll try to play all three games, and we’ll see whether we get anything playable at all. Hi, I’m Nick, and I’ve been developing software for over 20 years. On this channel, I share my experience, insights, and thoughts about IT. To keep the comparison fair, I prepared a minimal project with 3.js, TypeScript, and Vite. Rendering already works here, but there’s nothing else. No track, no car, no game logic. The task is to build a third-person arcade racing

01:35 game. It needs a proper closed-loop track, a drivable car, obstacles, collisions, checkpoints, three laps, and a finish. And all of it has to actually work, not just look nice in a screenshot. At the same time, I deliberately didn’t specify a visual style. I want to see what design and technical decisions the models make on their own. All three models work through open code with the same tools. Each one has its own copy of the project. They have 1 hour to write the code, run the

02:06 game, test it, and fix the bugs. Once they start, I don’t step in. Let’s start with Kimi. Generating a lot of code isn’t enough here. The model has to bring together car handling, track geometry, collisions, the camera, and the game logic, and all of that without a ready-made game engine. So, Kimi is done. It finished in 18 minutes, which is pretty fast. I expected it to take much longer. But, okay. What matters is that everything works. Let’s run it and see what we’ve got. All right, it looks

02:37 pretty good. A cyberpunk style. Just a reminder, in the task, I deliberately didn’t say anything about the style or what the game should look like. I wanted to see what the models would choose on their own. Well, functionality seems to be fine. The car drives, turns, collides with obstacles, but it looks like there’s a small problem with the obstacles after all. If an obstacle is slightly off to the side, you can drive right through it. But, overall, pretty good. I’d say the model handled the task successfully. Now, Claude Opus 5.5. I

03:09 take the same starter project and the same task. I launch open code, switch the model to Opus, and ask it to build the project the same way. The conditions are exactly the same. I’ll skip the generation itself, since nothing interesting happens there, and go straight to the result. So, Opus picked almost the same visual style as Kimi. Interestingly, both models named their games with neon. Kimi’s is Neon Ridge Raceway, and Opus’s is Neon Drift. But, in my opinion, the Opus game looks a bit

03:41 nicer than Kimi’s, though, of course, that’s subjective. No obvious problems, everything works as it should. To me, this version feels more like a finished game. By the way, Opus spent twice as long building the game, almost 40 minutes. Now, let’s take a look at GPT-6 version of the game. Oh, this is a different style. It looks great, but there are problems with the controls. The steering is reversed. When you press the right arrow, the car turns left. And when you press the left arrow, it turns

04:12 right. Very strange, but otherwise, it seems to work correctly. I looked into the code, and the cause turned out to be very simple. The right arrow increases the car’s heading angle, but in 3.js, a positive angle turns it left. So, the whole problem comes down to a single flipped sign. GPT did write tests for the game, but they only check the race rules, checkpoints, and laps. The controls weren’t tested, so the bug went unnoticed. If it weren’t for that steering bug, this would have been a

04:43 pretty decent game. So, what do we have after today’s game experiment? All three models got the same task, and each had the chance to make its own decisions, write code, and fix bugs. You could say all of them handled the task, except that GPT-6 had that strange steering bug. But what I was really interested in was how well the new Gemini K 3.1 Pro can compete at building 3D with the leading models from Anthropic and OpenAI. And it turns out that in

05:14 principle, it can. Visually, in my opinion, it’s weaker, but otherwise, it’s very good. On top of that, unlike GPT-6, the new Gemini model built a game without bugs that you could play right away. But it’s too early to draw final conclusions. I’ll keep using this model, but now for my everyday tasks, and I think in a few weeks, I’ll record another video with the results of using Gemini K 3.1 Pro over the long term. So, don’t forget to subscribe to the channel

05:45 so you don’t miss that video. Also, subscribe to my weekly newsletter about IT and AI. The link is in the description. See you soon. Take care.