Kimi K3 Develops Space Strategy Game in 6 Hours with Single Prompt… Comparative Test with Claude Fable 5, GPT-5.6 Sol Conducted

This game is the result of directly implementing all elements, including graphics, through code. It demonstrated the AI’s ability to plan and develop complex games autonomously without user intervention. In this process, Kimi K3 was developed using Three.js shapes, canvas, shaders, and web audio, without utilizing 3D models, textures, fonts, or sound source files.

Kimi K3 proceeded with game development by assigning 13 sub-agent roles. It repeatedly verified builds and actual execution screens, completing the game in 5 hours and 45 minutes. The Kimi K3 model, in particular, demonstrated strengths in long-term coding tasks and showed stable operation. In addition to the main agent, a separate critic agent was created to autonomously perform criticism and verification processes.

Furthermore, it supports a context window of 1 million tokens, giving it the ability to understand and process very long contexts.

The task involved providing only event brief information and asking for natural Korean copy, a poster studio, and Korean jamo animation. All three models successfully implemented the jamo animation. Each model was evaluated on its ability to design posters and write Korean copy using only the provided information.

This demonstrates differences in how each model interprets and visualizes information.

When comparing the Korean copy written by each model, the copy 'Moonlit Hanok Stroll' was evaluated as the most natural and appealing. In an internal AI evaluation, this copy written by the Kimi K3 model ranked first. The copy from GPT-5.6 Sol and Fable 5 models ranked second and third, respectively. This suggests that Kimi K3 demonstrates superior performance in Korean context understanding and creative writing.

All three projects passed all functional tests, including PNG export, adding copy, and jamo conditions. In the internal AI scoring, Kimi K3 received 96 points, Fable 5 94 points, and GPT-5.6 Sol 92 points, with Kimi K3 achieving a slightly higher score. This indicates that Kimi K3 demonstrated excellent performance in overall task execution, along with functional implementation.

This shows that Kimi K3 generated high-quality results but experienced some delays in terms of working speed compared to other models.

When converting token usage to API costs for the first task, the Kimi K3 model was the cheapest, costing $3.36. The Fable 5 model cost $14, and the GPT-5.6 Sol model cost $4.26. This indicates that Kimi K3 provided high scores along with cost-effectiveness.

This result demonstrates Kimi K3's high competitiveness in terms of cost efficiency.

It involved creating roads without external 3D models, allowing players to attach and swing on wires, and using inertia to move to the next building. Players had to pass five checkpoints in order and be able to return to the last checkpoint upon falling. This task demanded complex development capabilities, including 3D environment implementation, physics engine application, and game logic design.

While there wasn't much difference in the time it took for the first screen to appear for all three models, there was a significant difference in the total feature completion time. The Fable 5 model completed all features fastest in 31 minutes, followed by the GPT-5.6 Sol model in 51 minutes and 14 seconds. These two models perfectly implemented and verified all features within the 90-minute time limit. In contrast, the Kimi K3 model passed all features but exceeded the time limit, taking 2 hours, 27 minutes, and 24 seconds to complete.

It continued working for approximately 2 hours and 16 minutes after the time limit, passing all gates and creating the most spectacular and impressive result. Assuming no limitations, the Kimi K3 model received the highest score in technical evaluation. This case demonstrates Kimi K3's long-term autonomous learning and problem-solving abilities.

The GPT-5.6 Sol model cost about $8.4 using 10 million tokens. Kimi K3, based on the final completed version after exceeding the time limit, used 18 million tokens and cost about $9.06, making it the second most expensive after GPT-5.6 Sol. Kimi K3 used approximately twice as many tokens as Fable 5, but its cost was about half.

This was a process to test the AI's ability to distribute and perform complex large-scale projects.

Subsequently, it divided the work into Camera, World, Ship, Combat FX, Game Logic, and UI domains, and requested parallel execution from each sub-agent. A total of 13 sub-agents were configured, with 6 used for initial implementation. This demonstrates Kimi K3's ability to systematically partition and parallel-process complex projects.

There was no developer intervention during this process. All elements within the game, such as ships, planets, and resources, were self-implemented based on Three.js shapes, canvas, shaders, and web audio, without using external 3D models, textures, fonts, or sound source files. This result proves Kimi K3's complete autonomous development capability.

The fact that Kimi K3, an open model, could produce such a result, even when not a leading model like Claude or GPT, signifies tremendous progress in open model technology. Kimi K3 also showed the ability to autonomously delegate tasks to sub-agents during such long-duration operations. This is evaluated as a good example of Kimi K3's long-term autonomous execution capabilities and complex problem-solving skills.

Based on scores after the time limit was removed, the Kimi K3 model could achieve higher scores. This shows that Kimi K3 can perform complex tasks of very high quality if given sufficient time.

It used a main agent and 13 sub-agents to complete the task, and a separate critic agent was created to autonomously critique and verify the process. This demonstrates Kimi K3's very impressive ability to autonomously plan and execute complex, multi-stage projects over long periods, and to self-verify their quality.