Qwen 3.8 Flash, Mimo V2.6 Pro, and DeepSeek V4 Flash: comparing cost and quality in the generation of 3D scenes and games

A comparison of Qwen 3.8 Flash, Mimo V2.6 Pro, and DeepSeek V4 Flash shows how the three models handled generating 3D scenes and a game prototype in the same practical scenario. The author ran them through OpenRouter, used separate folders, and supplied identical prompts for three tasks: a low-poly fisherman, a Japanese pond, and a fishing simulator.

This is not an official benchmark or an independent study.

Below are the experiment author's observations, costs, and subjective assessment of the result quality.

What exactly was compared

The author tested not only the models' ability to write code, but also their ability to bring a project to a runnable state. For each model, the highest available reasoning level was selected, after which the projects were created in separate folders. This approach made it possible to compare costs and results within the same set of tasks, although it does not rule out the influence of the execution environment, random errors, and the specifics of a particular response.

OpenRouter and AI coding tools were used in the work. Qwen 3.8 Flash worked unstably for the author in OpenCode and constantly produced errors, so part of the test with this model had to be run in Pi. This limitation is important when interpreting the result: the observation applies to a specific combination of model and environment, not to all possible Qwen integrations.

Model costs and comparison limitations

The author provided pricing figures per one million tokens, but the original attribution of individual prices and release dates to specific model names was not sufficiently clear. Such information cannot be confidently used as a current price list for Qwen 3.8 Flash, Mimo V2.6 Pro, or DeepSeek V4 Flash without additional verification with OpenRouter and the model providers.

The table shows not prices, but the costs the author recorded across the three test tasks. The amounts are rounded and reflect a specific run, including the volume of reasoning, corrections, and repeated model actions.

The total was obtained by simply adding the rounded costs and is intended only as a rough guide. It does not replace a calculation based on current pricing: the actual cost depends on input and output tokens, repeated requests, and the terms of the specific provider.

Test 1. Low-poly fisherman

In the first task, DeepSeek V4 Flash was the cheapest option: the author reported a cost of about $0,08. Mimo V2.6 Pro required approximately $0,14, while Qwen 3.8 Flash, according to the author’s observation, cost approximately $0,28.

DeepSeek took first place for the visual result. The author found both the environment and the fisherman model itself convincing. However, operational stability became a problem: DeepSeek periodically stopped executing, and the author had to ask it to continue. The stated time of about five minutes was considered unreliable by the author.

Mimo produced an acceptable character with a fishing rod, bobber, plants, and a small animation. The weak point was the lake, which, in the author’s assessment, looked insufficiently developed. The scene was present, but the environmental detail was inferior to DeepSeek’s result.

Qwen created a semitransparent model with visual artifacts. The author also noted the poor placement of the plants. Even if the scene code ran, the final image required substantial manual refinement, so the higher cost did not provide a comparable gain in quality.

Test 2. Japanese pond

For the Japanese pond task, DeepSeek V4 Flash and Mimo V2.6 Pro cost approximately $0,13 each. Qwen 3.8 Flash cost about $0,19. The difference between the models was smaller than in the fisherman test, so scene quality and animation performance became the deciding criteria.

According to the author’s assessment, DeepSeek once again produced the best result. The fish moved, and the footbridge looked cohesive. For a small demonstration scene, this is an important result: the model did not limit itself to a static collection of objects, but added elements that created a sense of a living environment.

Mimo’s version included cherry blossoms and fish, but the fish movement seemed insufficient to the author. The scene conveyed a Japanese aesthetic, but the animation worked less effectively than DeepSeek’s. Qwen also implemented swimming fish, but it encountered problems with the cherry blossoms and the placement of the tree trunk.

And what matters more for this type of prototype: a collection of decorative objects or a cohesive environment? In the experiment, the author prioritized scene cohesion and object movement, and therefore ranked DeepSeek above the other two options.

Test 3. Fishing simulator

The game prototype was the most revealing test because it required the models to provide not just a single scene, but a combination of mechanics, interface, sound, and game state. Here, the costs were distributed differently: DeepSeek spent about $0,32, Mimo about $0,07, and Qwen about $0,44.

Qwen implemented a notable number of features. In the game, players could catch perch and crucian carp, choose whether to “catch or release,” change bait, buy worms and maggots in a shop, view records, and adjust mouse sensitivity and graphics quality. However, the author rated the prototype 2 out of 10 because of unfinished elements and weak three-dimensional fish models.

Mimo’s low cost in this task did not mean that the game was ready for use. According to the author’s observation, unclear problems occurred at launch, and there was no sound. In the first version tested, the controls were so unclear that the author did not understand how to catch fish: the distance to them increased, while performing the action required clicking the mouse. The game included bait, a keepnet, and a shop, but the author called it completely buggy.

DeepSeek created the most cohesive of the three game prototypes. It included sound, a correctly moving bobber, line tension, and friction adjustment. The author also noted the acceptable quality of the fish and three-dimensional models. The player could sell caught fish, buy items, use bait, a keepnet, and fishing lines, keep records, and change the graphics quality.

Increasing the graphics quality in the DeepSeek version, according to the author’s observation, did not produce a noticeable visual improvement. This is a limitation of the prototype: having a setting in the interface does not necessarily mean that it actually changes rendering or asset quality.

Why Qwen’s result cannot be generalized to all execution environments

Qwen 3.8 Flash constantly produced errors for the author in OpenCode, so the test had to be continued in Pi. This change of environment makes the comparison less clean: the model may have encountered problems with integration, tools, file operations, or a specific execution mode, rather than only its own limitations.

The same applies to DeepSeek’s instability in the first test. The author observed execution stoppages and had to ask it to continue, but a single scenario is insufficient to declare the model unstable in all environments. For a practical choice, a developer should separately test launching, recovery after errors, file handling, and the model’s ability to complete a long task independently.

Comparing cost and results

If only individual costs are considered, Mimo was the cheapest in the game test, while DeepSeek spent more. The author compared not the price of a single run, but the balance between cost and the usefulness of the result. In the author’s experiment, DeepSeek delivered the most playable version, so the additional expense seemed justified.

• DeepSeek V4 Flash performed best in the cohesion of 3D scenes, fish movement, and the creation of a game prototype with basic fishing mechanics.
• Mimo V2.6 Pro produced acceptable results in the fisherman and pond scenes, but the game launch was problematic, and there was no sound.
• Qwen 3.8 Flash created a game with a large number of stated features, but the implementation quality, visual models, speed, and operational stability received low marks from the author.

Is the lowest price worthwhile if the result requires a complete rebuild? In this test, the author’s answer was no. For quick visual prototypes, the author chose DeepSeek despite its failures and its not-lowest cost in the game task.

The author’s final ranking

1. DeepSeek V4 Flash. First place for the most interesting result when quality and cost are considered together. The model created the best, in the author’s opinion, low-poly fisherman, the most successful Japanese pond, and the most playable fishing simulator.
2. Mimo V2.6 Pro. Second place. The scenes were sufficiently acceptable, but the lake was weak, the fish movement was limited, and the game version encountered launch problems and lacked sound.
3. Qwen 3.8 Flash. Last place in this particular experiment. The author noted slow performance, a large number of errors, artifacts in 3D scenes, and a mediocre game result despite the presence of several game mechanics.

Comparison limitations

The ranking cannot be considered a universal conclusion about the capabilities of the three models. The author tested a small set of tasks, did not conduct independent repetitions, and assessed visual quality subjectively. The result may have been affected by the tool version, execution environment, file operation errors, differences in the number of iterations, and the need to manually ask the model to continue working.

The prices also apply to specific runs through OpenRouter rather than fixed model tariffs. Before choosing a model for ongoing development, it is worth separately checking the current token cost, the model’s availability from the required provider, integration stability, and the need for manual code correction.

Within the scope of the described test, DeepSeek V4 Flash appears to be the most practical choice for generating simple 3D scenes and a game prototype. Mimo V2.6 Pro is suitable for inexpensive environment experiments, while Qwen 3.8 Flash requires additional testing in another environment and stricter quality control before being used for similar tasks.