DeepSeek V4.1 Flash and Four AI Harnesses: Comparing Game Character and Animation Generation

The article examines a user test of DeepSeek V4.1 Flash in four AI harnesses. This term refers to software shells that manage an agent, its tools, files, and work session. The author compared P Agent, Claude Code, OpenCode, and DeepSeek’s native harness by the quality of game assets, token usage, cost, and working time.

The result does not claim to be a universal ranking of models or a definitive benchmark. It is the author’s observation from a single scenario, supplemented by a subjective five-point rating. The outcome was influenced by the settings, context, order of refinements, and specifics of the particular session.

Harness Comparison Methodology

The test followed a sequential production scenario. First, each harness had to create a basic orc, followed by a set of enemies: a simple warrior, a heavy warrior, an archer, and a shaman. After that, the author tested dioramas and small games assembled from the created characters, weapons, and animations.

At each stage, the author recorded input and output tokens, cache usage, money, and time. Quality was assessed on a five-point scale, checking not only appearance but also the result’s suitability for a game. Important indicators included the correct positioning of hands, fingers, and weapons, silhouette readability, animation performance, enemy behavior, and game balance.

A single image rating reveals little here. A character may look convincing in a static pose, while movement exposes problems: the hilt passes through the palm, the bowstring does not align with the hands, or enemies collide with one another. For game assets, defects like these determine the amount of subsequent manual work.

The Basic Orc in Four Harnesses

The first orc from P Agent took 32 minutes 7 seconds to create, required about 296 thousand output tokens, and cost approximately 36–37 cents. The author did not specify input usage in this part of the test. The silhouette reads as an orc, with clothing, tusks, nails, and pronounced musculature. The author noted a strange face, slanted pupils, tucked thumbs, and four toes on each foot. The running animation looked unnatural: the palms were turned upward.

Claude Code spent 59 minutes 9 seconds on the orc and used about 49 cents. Generation volume amounted to approximately 429 thousand output tokens. The face turned out meaner, the tusks looked more reasonable, and the silhouette remained recognizable. However, the face seemed fused into a single mass, the thumbs were tucked again, and the author counted four toes on each foot. During running, the palms also pointed upward.

OpenCode used approximately 153–154 thousand input tokens, 69 thousand output tokens, and 36 million cached tokens. The cost was approximately 31 cents, and the work took 51 minutes. In the author’s assessment, this orc turned out slightly better than the previous two: the silhouette was more accurate, facial features were clearer, and the musculature was well developed. The main defect concerned the costume, whose rear section visibly stood away from the body. The author considered this problem easy to fix.

Among the variants shown, OpenCode’s running animation received the highest rating. The palms pointed downward, and the gait looked natural enough. The model not only created the outward appearance but also better connected the position of the hands with the character’s movement.

DeepSeek’s native harness used approximately 116 thousand input tokens, 275 thousand output tokens, and about 27 million cached tokens. The cost was about 39 cents, and the work took 31 minutes 24 seconds. The transcript also contained the value “20” next to the cache, but the author clarified the final value as 27 million.

DeepSeek’s orc had well-developed musculature, a face, and body details. The author described the expression as cunning rather than foolish or angry. The shortcomings included tusks that were too large and poorly attached, as well as an incorrect number of fingers: according to the author, each hand had four fingers, while each foot had three toes. The running animation again looked unsatisfactory because the palms were turned upward.

P Agent received three stars for its short working time and low costs, along with a sufficiently detailed appearance. Claude Code received two stars because of its long working time, higher cost, and mediocre result. OpenCode created a slightly higher-quality orc, but the author lowered its rating because of the duration and defects.

Refining the Orc: Armor, Weapons, and Attacks

To refine the orc, P Agent used 131 thousand input tokens and 226 thousand output tokens, about 24 million cached tokens, 34,6 cents, and 26,5 minutes. The harness added spiked armor and an axe over the previous clothing. The model positioned the weapon incorrectly: the hilt passed through the hand.

The movement set included a regular strike resembling a golf swing, a side strike, and a battle cry. The side attack looked more convincing than the basic strike. This stage clearly showed the limits of automated generation: an outer equipment layer can be created quickly, but the weapon’s contact with the hand requires checking and manual correction.

Heavy Warrior, Archer, and Shaman

Claude Code created the heavy warrior using 460 thousand input tokens and 497 thousand output tokens. The cache amounted to 13 million tokens, the cost reached 52 cents, and the work took 48 minutes 24 seconds. The heavy armor looked visually successful, but the two-handed axe was held incorrectly: the arms were twisted, and the grip looked unnatural.

The heavy warrior’s striking motion itself looked as expected: the character struck from above. The set also included a spinning attack and a battle cry. The arm defect remained, so the attractive costume did not compensate for the error in the basic weapon-interaction mechanics.

OpenCode spent 129 thousand input tokens and almost 312 thousand output tokens on the archer, along with about 26 million cached tokens, 28,5 cents, and 35 minutes 14 seconds. The harness added a hood or cloak, a quiver, arrows, a bow, and additional armor. The bow and arrows passed through the hands and were held incorrectly.

The shooting animation worked partially. The bowstring drew back, the bow fired, and the arrow flew into the distance, but the motion itself looked strange. The set also included a double shot and a battle cry.

OpenCode’s shaman required 119 thousand input tokens and 283 thousand output tokens, almost 22 million cached tokens, 36,5 cents, and 31 minutes 32 seconds. In the author’s opinion, this was the most interesting appearance: a glowing bone staff with a crystal was held relatively correctly, bone elements appeared in the clothing, and the eyes received symbolic changes.

The shaman’s animations included throwing magic with the hands, striking with lightning from the staff, and a battle cry with an additional ground effect. Despite the initial problems with the number of fingers, the author called this generation one of the most interesting. In the final refinement rating, the shaman received four stars as the cheapest, fastest, and highest-quality result among the enemies considered.

Dioramas: Scene Quality Versus Time

P Agent’s diorama cost 76,6 cents and took 48 minutes 38 seconds. The scene used almost 209 thousand input and 339 thousand output tokens, while the cache amounted to nearly a million tokens. The scene included the previously created orcs, a camp with a fire, houses or warehouses, a forge, watchtowers, and grazing animals. The author found only a few noticeable defects, including the strange movement of the orc with the two-handed axe and some character collisions.

Claude Code’s diorama cost approximately 45 cents and took 52 minutes 28 seconds. It used almost 372 thousand input and 374 thousand output tokens, while the cache amounted to about 17,45 million tokens. The author immediately noticed an upside-down roof, an incomplete warehouse, and animals that looked more like dogs than boars.

OpenCode spent 53 cents and 83 minutes 56 seconds. The volume amounted to 631 thousand input and 214,5 thousand output tokens, while the cache reached 103 million. Strengths included lighting, shadows, fires, and lights on the watchtowers. Weaknesses appeared in the behavior of the orcs, which collided, stopped, and stumbled over one another, as well as in the geometry of the warehouses, roofs, cart, and palisade.

For DeepSeek’s native harness, the record lists 248 thousand input, 441 thousand output, and 88,8 million cached tokens. The cost was about 96 cents, and the work took 74 minutes 53 seconds. The author considered the scene the most adequate in terms of overall assembly and lighting: the wheels were installed correctly, and sparks, a flag, a forge, a table, and elements resembling chairs appeared.

This diorama also had limitations. An overexposed circle noticeably changed the shadows as it rotated, the orcs continued to collide, and some character behavior remained unpredictable. The neural network assembled the final scene from elements of different generations rather than the author assembling it manually. As a result, the table was disassembled, and some defects remained.

Game Prototypes and Practical Usability

P Agent created a game in 51 minute 20 seconds. Input usage amounted to 187 thousand tokens, output usage to 253–254 thousand, the cache to about 3,637 million tokens, and the cost to about 46 cents. Controls used WASD, attacks were performed with the mouse, and the mouse also handled viewing and interaction.

The author liked the game itself, but the balance was harsh: the hero had little health, there were many enemies, and opponents used ranged attacks. Without a block and healing items, defeating all the orcs was practically impossible. The shaman usually retreated and did not attack the hero, but the author did not confirm the assumption that the shaman had special protection.

Claude Code’s game took about 97 minutes and cost approximately 99 cents. The author reported about a million input tokens and 72 thousand output tokens, but some cache data was illegible. The character and animations looked somewhat worse than in the comparison variant, although combat was slightly easier. This was the only game the author managed to complete, taking the gold.

OpenCode ran for about 138 minutes, used 211 thousand input and 390 thousand output tokens, and had a cache of 84 million. Controls were performed with the left mouse button or the spacebar. The sword strike looked more interesting, but the game periodically froze. The archer and mage almost never missed the moving character, while blood from hits remained suspended in the air.

The game from DeepSeek’s native harness took about 74 minutes 52 seconds, cost approximately one dollar, and required 219 thousand input, 396 thousand output, and 99 million cached tokens. The character looked more angular, but the strikes seemed adequate. Inverted vertical controls, a dense group of enemies, and no healing made the game extremely difficult to complete.

The Author’s Final Ranking

1. DeepSeek’s native harness. It was the most expensive, but worked relatively quickly and produced stable, generally good generations.
2. P Agent. In most cases, it worked faster than the others and spent less money and time, but quality varied noticeably.
3. OpenCode. The harness created interesting variants and showed its strongest performance in animations, but worked the longest and retained errors in hands, weapons, and object behavior.
4. Claude Code. It took last place because of its lengthy runtime, high cost, and mediocre results, although its game had the best balance and was the only one that could be completed.

The test’s main conclusion concerns not the harness’s abstract “strength,” but its suitability for the task. For rapid prototyping with spending under control, P Agent looks more appealing. For more stable, comprehensive generation, the author chose DeepSeek’s native harness. OpenCode is interesting when individual successful animations and visual solutions matter, but its long sessions require patience. Claude Code did not lead in assets, but it produced the most playable game scenario.

The five-point and star ratings reflect the author’s opinion, not an objective quality standard. Before choosing a tool, a developer should repeat the test on their own project and separately check the hands, fingers, weapon attachment points, collisions, enemy navigation, controls, and balance. These checks are what show how much manual work will remain after generation.