Kimi K 3.1 Pro in practice: testing a 3D orc, an aquarium, a game, and website design

Kimi K 3.1 Pro is a multimodal model for working with text, images, video, and code. In the author's practical tests, it created a 3D scene with an orc, an interactive aquarium, a Super Mario-style browser game, and a studio website design. The results were uneven: the model handled logic and frontend development well, but did not always bring visual details to a final-quality standard.

All evaluations and results below refer to the author's observations in specific test scenarios. They do not replace a direct comparison of models using identical prompts, identical limits, and a unified set of criteria.

What is known about Kimi K 3.1 Pro

According to the author, at the time of testing, Kimi K 3.1 Pro supported a context of up to 1 million tokens and accepted text, images, and video. The author also stated the listed price: 2 dollars per 1 million input tokens and 10 dollars per 1 million output tokens. These terms refer to the data available to the author during the experiment. They should be checked separately before using the API or a third-party interface.

In the tests, the model was used not only as a chat tool for code generation, but also as an executor of multistep tasks. It was asked to create scenes, script object movement, add sounds, configure interfaces, and fix errors after the user interacted with the result. This format makes it possible to evaluate not just an individual code fragment, but the model's ability to maintain a project's structure throughout a lengthy session.

Methodology and interfaces

For the aquarium, the author chose Open Code because the interface displayed expenses, execution time, and token usage. These metrics were not shown in the Kimi interface being used, so the costs of different tests cannot be compared directly without qualification. According to the author, Kimi K 3.1 Pro was available to users with the relevant Open Code Gold subscription.

The author launched the game project with the highest reasoning level, using Open Code Go and Kimi K 3.1 Pro. After exhausting the Open Code Go limits, he continued working through OpenRouter. The final game reflects the operation of a combination of interfaces and providers, not just one model in an isolated environment.

3D orc: strong form and successful animation

In the first test, Kimi K 3.1 Pro created a low-poly orc model with a recognizable face, armor, scars, and a sharpened axe. The author considered the design sufficiently detailed: the character looked cohesive, and the key features of an orc were clear without additional explanation. For a game prototype, the result was usable after the first pass.

The shortcomings appeared in the smaller elements. The author noticed errors in the fasteners, belt, knees, and boots, while the hands were less detailed than the other body parts. These defects did not destroy the overall silhouette, but the model would require manual correction before being used in a full scene, especially if the character were to be shown in close-up or take part in complex animation.

The animation turned out better than the geometric details. The orc raised his arm, swung the axe sharply, and struck downward, after which a striking sound played. In the author's assessment, this was one of the best animations among the models he had previously tested. This assessment applies to the specific experiment and does not represent a universal ranking of Kimi's capabilities.

The author suggested that the orc's quality was comparable to the results of GPT Soul and Astra on similar tasks, but he did not conduct a direct comparison. Differences in prompts, environments, number of iterations, and available tools prevent this result from being considered a full benchmark.

Aquarium: object logic matters more than visual polish

Creating the aquarium took the author about 45 minutes, and the stated cost of the experiment was 3 dollars with approximately 140 000 tokens used. Marine creatures moved in the scene, an aerator operated, and bubbles created visual distortion. The author assessed the aquarium as a normal working result, but did not see a notable advantage over the work of Saneta or Haiku.

The crab moved sideways while keeping its claws intact, and the snails had eyes, a recognizable shape, and crawled around the scene. These details show that the model connected an object's appearance with basic movement logic, although the behavior of different elements remained unevenly stable.

The author called the plants ordinary and insufficiently detailed. The aerator effect worked, but the wire was severed in one place. One fish remained motionless, and the author could not determine whether this was related to the specified movement logic or to collisions. The fish did not create noticeable turbulence in the water, but the bubbles correctly produced visual distortion, and the author considered the wave behavior logical for small bubbles.

So what does this test show? Kimi can assemble a scene in which individual objects interact and move, but visual detail remains uneven. That is enough for a prototype or demonstration. A commercial 3D scene would require collision testing, animation review, replacement of weak models, and manual material adjustments.

A Super Mario-style browser game

Game generation took about an hour and cost 6,32 dollars according to the author's estimate. The project included a starting window with instructions, music, enemies, character controls, a background, a growth mushroom, destructible blocks, coins, an underground level, a shooting flower, and a star. This set of mechanics made it possible to test not only the appearance, but also whether the world state was preserved after player actions.

The character looked good enough, but walked sideways. Enemies could be defeated by jumping, the flower allowed the character to shoot, and the world persisted after the character's death. One of the enemies did not stop, and the author was unable to eliminate this behavior during the test. The model also correctly restored an object after interaction and generally maintained coherent game logic.

The author's statement about the “perfect” construction of the logic expresses his assessment of a specific result, not a formal quality metric. The game included a castle, a door the character could enter, fireworks, secret elements, and a hand-raising animation. The level ended after the time expired, indicating the presence of a scripted limitation or completion timer.

The game's interface visually resembled projects such as Mario and received a fairly high evaluation from the author. This result is useful for rapid prototyping: the model can assemble a start screen, game elements, and several interconnected mechanics in one project. Before publication, controls on different devices, collisions, balance, sound, interface accessibility, and correct level completion would need to be checked.

Studio website design

Creating the vibe-coding studio website took 33 minutes 10 seconds, and cost nearly 2 dollars. Kimi created a page with a 3D object, a heading, block navigation, and interactive project elements. The contact form included a request-type selector, fields for email and name, and a field for describing the idea.

The author rated the design as average and fairly decent. The website did not look obviously bad or entirely template-based, but individual sections, the 3D element, and the project cards required further work. The responsive mode worked correctly, so the author rated Kimi highly specifically as a frontend developer and executor of visual tasks.

The result was less convincing when it came to generating an original concept. The author gave the model the subjective score “50 out of 50”: it followed the assigned instructions well, but did not seem to be as strong a source of unusual ideas. For a designer, this means preparing the visual direction, references, page structure, and stylistic constraints in advance, and only then asking the model to assemble the interface.

Comparison with other models

How much the practical experiments cost

Three expense estimates are known from the scenarios described: the aquarium cost about 3 dollars, the game required 6,32 dollars, and the website cost nearly 2 dollars. The creation of the 3D orc has no separate cost listed in the source data. The total amounts cannot be extrapolated to every project: usage depends on the length of the conversation, number of corrections, reasoning mode, selected interface, and provider pricing.

• Aquarium: about 45 minutes, 3 dollars, and approximately 140 000 tokens.
• Game: about an hour and 6,32 dollars.
• Website: 33 minutes 10 seconds and nearly 2 dollars.
• 3D orc: no separate cost is specified in the test description.

The author considers Kimi a fairly expensive model and notes that cheaper alternatives are available, although no precise cost difference is provided. Before launching regular development, it would be sensible to conduct a short test of your own: use one prompt, limit the number of iterations, record token usage, and separately assess the time required to fix errors.

Who Kimi K 3.1 Pro is suitable for

Kimi K 3.1 Pro appears to be a suitable tool for developers who need to quickly assemble prototypes of interactive websites, browser games, and multimedia scenes. The model is particularly interesting for tasks in which code, visual elements, animation, and user interaction need to be connected in a single workflow.

For final design, complex 3D graphics, and projects with critical stability requirements, a single generation is not enough. Errors in a character's hands, the aerator wire, the fish's movement, the hero's walk, or the enemy's behavior show that the result must be tested after every major change rather than accepted based on one successful screen.

Final assessment

According to the author's observations, Kimi K 3.1 Pro showed an average overall result among strong models, but performed particularly convincingly in frontend development and game-logic construction. It managed to create a recognizable 3D orc, assemble an aquarium with moving objects, implement a game with several mechanics, and build a responsive website with a contact form.

The main limitation is the unevenness of the quality. The model can quickly assemble a cohesive prototype, but small details, collisions, character movement, and the originality of the visual concept require checking and manual refinement. A direct comparison with GPT Soul, Astra, Saneta, Haiku, and Opus Fable has not yet been conducted, so the current conclusions should be viewed as the author's assessment of a practical experiment rather than a definitive ranking of models.

When choosing Kimi K 3.1 Pro, not only the stated price and context size matter, but also the specific workflow. If the priority is fast vibe coding, interface assembly, and game prototyping, the model looks promising. If a predictable production-ready result without manual oversight is required, the tests indicate the need for additional engineering verification.