DeepSeek V4.1 Flash vs. GLM 5.3 Flash and Gemini 3.8 Flash: a test of 3D scenes and games in vibe coding

The comparison of DeepSeek V4.1 Flash, GLM 5.3 Flash, and Gemini 3.8 Flash is based on a practical test described by the reviewer. The author tested the generation of 3D scenes and game prototypes through AI agents, and then evaluated the result, operating speed, and approximate costs. This is not a laboratory benchmark or a universal ranking of the models.

The author named DeepSeek the best.

The results and conclusions of this particular test are presented below. The primary quotation from the original video is unavailable in the materials, so the wording should not be regarded as independent confirmation of the models’ characteristics.

What was compared and how the test was conducted

The models were tested on three types of tasks: creating a 3D model of a homeless person, assembling a small diorama of a landfill, and developing a game in which a character fends off opponents with garbage and bottles. A high level of reasoning was enabled for all three models. According to the test design, this mode increases the task completion time and makes it possible to evaluate not only the writing of an individual piece of code, but also the agent's work with a large project, scene files, interaction logic, and bug fixing.

DeepSeek V4.1 Flash 1017 was run in OpenCode using original API access. GLM 5.3 Flash was tested through an OpenCode Go subscription, while Gemini 3.8 Flash was run in Antigravity with an original subscription. The access environments differed, so the final cost in this test depended not only on the model, but also on the plan, limits, billing method, and how the specific interface counted requests.

Gemini's absolute cost cannot be directly compared with DeepSeek's and GLM's expenses. The author did not record the exact costs and duration of individual tasks for Gemini, instead estimating usage based on the share of the available limit. Approximate hourly and weekly expenditure figures appeared in different parts of the test, and one of the early estimates was later acknowledged to be inaccurate.

Test of the 3D model of a person

DeepSeek completed the task in 33 minutes and 30 seconds, with a recorded cost of 30 cents. GLM finished the work in approximately 15 minutes and spent 12 cents. The exact time and cost for Gemini in this episode were not established.

Gemini created a model with a square shape. The character had trouble with one leg, and the animation was limited to slight body swaying and a small amount of scarf movement. The author found this result sufficient to demonstrate the idea, but in their assessment, the model looked more like a rough draft than a finished game asset.

GLM produced a slightly better visual result. The legs looked better, but the hand remained problematic, the shapes were excessively angular, the detail was low, and the nose and certain facial features had noticeable artifacts. The character also swayed slightly, so within the scope of the test, it could not be called completely static.

According to the author, DeepSeek created the most refined model of the three options. The fingers and thumb were distinguishable, and coins were visible in the glass. Errors remained: one hand was poorly executed, and the glass ended up inverted or poorly assembled. The level of detail was higher, but the developer would still have had to manually check the geometry and fix individual parts.

But what should you do if Blender produces an unstable result? The author believes that directly creating a model in Blender is not suitable for every scenario he tested. In his experience, it is sometimes more sensible to first obtain an object in 3GS and then ask the agent to transfer it to Blender or prepare it directly for a game. This is the author's practical recommendation, not a universal rule for all projects and models.

Junkyard diorama: detail density and scene realism

The second test examined how the models assemble not just a single object, but an entire scene with an environment, background, garbage, and simple animations. DeepSeek completed the task in approximately 40 minutes and spent 25 cents. GLM worked for about 48 minutes, and the cost was 62 cents. The exact cost for Gemini was not confirmed, since the author estimated it based on the proportion of the limit used.

Gemini added water, rocks, a fence, a refrigerator, a sofa, garbage bags, and wheels. The scene contained the expected elements, but some objects were translucent, there were few details, and the author rated its resemblance to a real junkyard as moderate. In his view, the individual objects did not make up for the weak composition and insufficiently filled space.

GLM assembled a more convincing junkyard. The scene featured a pile of garbage, a bulldozer, a sofa, a refrigerator, water with droplets, a floating leaf, and moving grass. According to the author, this diorama conveyed the required setting more effectively than Gemini's result because the model added not only static objects but also several environmental elements.

DeepSeek created the most richly detailed scene. It contained a diverse pile of garbage, bottles, boxes, barrels, a container, a refrigerator with a door, tires, water, and a floating leaf. The author preferred this result for dioramas because the scene better matched the original task and provided more material for further development of a game level.

In the scene test, the results demonstrated a limitation of the approach the author calls vibe coding: an agent can complete a list of objects, but it does not always distribute them correctly in space, maintain scale, or prevent intersections. Translucent surfaces, incorrect materials, and overly simple geometry require a separate check in the editor, even when the original prompt was detailed.

Game Testing: Mechanics Matter More Than the Initial Visual Shell

In the game test, GLM finished in approximately 6 minutes and spent 8 cents. DeepSeek took about 41 minutes, while the cost came to 32 cents. The exact cost of the Gemini result was not established. GLM was noticeably faster, but within this comparison, generation speed alone did not demonstrate that the game was ready for use.

One of the games created by GLM had sounds and the ability to throw objects. The player could throw trash and bottles, and the enemies also reacted by throwing objects, but the “hit” and “search” actions did not work, and the R key did not respond. The author called this version semi-functional: the main idea was clear, but the key interactions remained broken.

In another tested game, attacks, knockouts, searching trash cans with animation, throwing trash, and turning in bottles all worked. The enemies could also throw objects at the player, and it was possible to complete the scene. The author rated the visual models as extremely weak, so they did not equate the functional logic with a full-fledged prototype.

Another game allowed the player to switch between bottles and trash. The enemies reacted to the player's actions, while picking up objects and searching were accompanied by animations. Searching was interrupted when an enemy threw an object, and one of the enemies could only attack by throwing objects and did not engage in melee combat.

The models, houses, and animations looked best in the final version. The boss had more health than the other enemies but moved slowly, so the author was able to defeat it by pelting it with trash. In the author's assessment, this was the only game that could already be used as a prototype for further development, although testing, optimization, and checks on different devices would still be required before publication.

Summary Comparison of the Models

In this test, DeepSeek received the author's higher rating for the level of detail in the 3D model, the richness of the diorama, and the quality of the final game prototype. GLM won on speed and proved to be the cheapest in the game test, but its fast result required checking nonfunctional actions and weak models. Gemini showed a more modest result in the scenes described, and its costs cannot be correctly compared with those of the other two models without precise data on the limits and pricing.

The author called DeepSeek the best option for long-term budget vibe coding. This conclusion applies to the scenarios described and is not an official assessment by the model's developer. The choice will change if the priority is the speed of short iterations, a specific editor integration, an available pricing plan, or a different type of project.

What to check before publishing a game on Yandex Games

The resulting prototype cannot immediately be considered ready for publication. Yandex Games' requirements specify correct SDK integration, the absence of technical errors, progress saving, functional controls on the stated platforms, and a finished-looking game. The game must not contain development traces, nonfunctional buttons, or critical errors in the browser console.

The service rules also contain requirements for the description to match the actual mechanics, the quality of promotional materials, and the correctness of advertising. The published Yandex Games materials state that full moderation usually takes 3–5 business days, while the wait in the queue depends on the workload and the account's history. For new accounts, the wait, according to the description provided in the materials, may take several weeks.

Rejection increases the overall publication timeline. After the first rejection, resubmission is available no earlier than after 24 hours; after the next rejection, the interval increases to 2 days, then to 4, 8, and 16 days; after the fifth rejection, the maximum interval is 384 hours. These intervals apply to resubmission after rejection and do not replace the time required for the review itself or the time spent waiting in the queue.

In the test account, the author estimated the overall time to publication at roughly three weeks and assumed an additional three weeks or so after an error or rejection. Official Yandex Games materials distinguish between the queue, the moderation itself, and the intervals between resubmissions, so the actual timeline depends on the specific account, workload, and number of revisions.

Conclusion

DeepSeek V4.1 Flash received the author's highest rating in the scenarios described, where details of the 3D scene, environmental richness, and gradual refinement of the game prototype are important. GLM 5.3 Flash is worth considering when the speed of the first version is critical and the user is prepared to check the mechanics manually. Gemini 3.8 Flash performed worse in terms of visual results in this comparison, while its cost remained uncertain.

The practical conclusion of the test applies to the scenarios examined: an AI agent helps quickly produce a foundation but does not replace checking the scene, gameplay actions, animations, and publication requirements. The more complex the project, the more important it is to separate the generation, testing, and correction stages rather than evaluate the model solely by how impressive the first launch looks.