A 3D action game made with 450 million tokens—can DeepSeek do it again?

Guizang posted a browser game called Mist-Locked Mountain Gate on X; click the link and you can play. WASD to move, left-click to attack with the sword, right-click to block, space to dodge. The file is a single HTML file, 30 MB. According to him, the model, scene design, action design, and optimization were all done by Fable 5.2 in a single afternoon. Playable link: https://op7418.github.io/guizang-game-skill/

In the post, he first wrote 4.5 billion tokens, then corrected it to 450 million in a reply. The optimization part is more specific: the original assets totaled more than 100 MB and were compressed to 30 MB with little loss in quality. The workflow used Tripo and the Jimeng CLI, letting the model generate concept art and 3D models itself, then do rigging. Videos in the replies show the action and scenes.

The process was not one-shot. He replied that he went through many revision rounds, drew on others' experience making games with AI, and also wrote a skill. The skill first clarifies the requirements, then researches tools, and after that drives the visuals, self-checks, and verification. The combat techniques, hit feel, and optimization for action games are all written into the skill, along with a list of likely pitfalls; one key focus is how to package large assets into a small build without affecting playability. GitHub link: https://github.com/op7418/guizang-game-skill

Then he asked: could his skill and DeepSeek Harness replicate Fable 5.2's success? DeepSeek Harness can be roughly understood as connecting the DeepSeek model to tools so it can run long tasks. The comments section started arguing right away. Michael Anti replied bluntly, 'Probably not,' and Guizang was just as direct, saying it would be tough. One commenter said DeepSeek's 3D capabilities were not optimized in this direction; someone suggested adding a control group by running the same skill on Fable again, otherwise it would be impossible to tell what deserved the credit; someone wondered whether the bill for 450 million tokens could be cut to the price of a hot pot meal; someone just wanted to see the failure step. Another commenter guessed it could get 70–80% of the way there, with the gap lying in the judgment calls during repeated rejections in the middle, which cannot be written into a skill.

Other replies also posted more browser games made with Fable 5.2, including shooters and racing games. This kind of attempt is not a single sample, but replication is not about one-off results; it is about whether it drifts off course during long tasks and whether it can recover on its own when it hits visual and feel problems.

Some people in the comments also stacked group members' avatars, making it feel like opening a betting pool. If it really works, the value is not only in saving on the bill but also in verifying how much experience a skill can carry over. If it does not work, it can still show which judgments still need to come from a human and which workflows can already be standardized. The DeepSeek Harness comparison is worth waiting for results on.