The comparison of Haiku 5.5, GPT-6 Luna, and DeepSeek V4.1 Flash is based on the author’s observations from several runs: creating a low-poly 3D character, a 2D platform fighter, and an interactive village scene.
The ranking is subjective. The author evaluated the projects after launching and visually inspecting them, so the conclusions cannot be applied to other development tasks.
What exactly was compared
In the tests, the models were given tasks involving the creation of game and visual scenes. The details of the run settings, including folder organization and reasoning parameters, are not confirmed by the available evidence.
The author compared three scenarios:
- a low-poly 3D character with a magical theme;
- a 2D platform fighting game with attacks, grabs, levels, and bosses, if these elements appeared in the generated result;
- an interactive village with houses and characters performing different actions.
For the village scene, they used a text prompt and a reference image showing the placement of buildings and characters. According to the author’s observation, all three models were able to interpret the image, but the quality of transferring the layout and the behavior of the villagers differed.
Test methodology and limitations
The author launched the created projects and evaluated them based on visual results, how the mechanics worked, and subjective impressions of the scene. The available evidence does not confirm a complete system for measuring generation time, tokens, limits, and cost.
The cost comparison is also limited. The evidence mentions a Claude subscription costing about 20 dollars, while for DeepSeek the author noted payment for individual requests. One DeepSeek run, according to the author, cost about 22 cents, while some requests reached 70 cents.
Tokens, subscription limits, and the price of an individual request describe different parameters. They do not provide a single measure of quality or efficiency, especially when the models operate through different access interfaces.
The model names in the original material were rendered inconsistently, including the variants Haku, Haiku, HKO, Deepsek, Deepsik, and DeepS V4.1 Flash. The article retains the names from the assignment: Haiku 5.5, GPT-6 Luna, and DeepSeek V4.1 Flash. Before repeating the test, a developer should separately verify the exact model identifiers and the availability of the required versions.
Test 1: low-poly 3D mage
According to the author’s observation, the DeepSeek result looked primitive. The character could not be freely rotated, the crystal was positioned incorrectly, the staff sat poorly in the hand, and the animation seemed strange.
Luna still had problems with the hat, but the character held the staff; the scene contained a rotating magical orb and other effects.
The author called the Haiku result the most developed. The character raised its head, the hat and staff looked more appropriate, the staff was held in the hand, and the scene included a glow, particles, staff effects, and elements on the floor.
The author assumed that Haiku’s longer processing time may have helped produce a better result. This is an assumption, not an established causal relationship: no separate measurement was conducted of the effect of generation time on quality.
Test 2: 2D platform fighting game
The Luna project launched, but the author did not find clear attack animations or working controls on the specified JKL keys. Visually, the game looked acceptable, but in the author’s assessment it was unfinished and practically unplayable as a fighting game.
Haiku displayed attacks, enemies, jumping attacks, a background with a bar, enemy counterattacks, checkpoints, a boss, and at least three levels.
The item pickup or grab mechanic assigned to the G key did not work, or was implemented so unclearly that the author could not use it. The other elements allowed the author to assess the project as a complete and generally decent prototype, although some features required manual verification.
In the DeepSeek version, item grabbing was assigned to the E key, but the author did not find any working interaction with the items.
The situation with enemies looked different: DeepSeek allowed the player to grab, hold, and throw them. The author also noted dodges, different types of attacks, grabs, and attacks using a grabbed opponent.
The boss could dodge, although some attacks, including a kick, made it possible to defeat it quickly. In the author’s subjective assessment, the DeepSeek fighting game was more interesting than the Haiku version in terms of scenery, movement, and character variety. The graphics resembled a LEGO fighting game, and the gameplay dynamics looked more convincing.
DeepSeek took first place in this category specifically according to the author’s assessment. The conclusion applies to this run and is not an independent measurement of model quality.
Test 3: interactive village
The village scene made it possible to compare the visual placement of buildings and characters, their actions, animations, and collisions. The complete list of methodological criteria was not separately documented in the evidence, so the observations below reflect the author’s observations of the results.
In the Haiku version, the fisherman looked engaged in fishing, the children were appropriately scaled to the houses and ran around the village, and the well visually resembled a well. The author found no collisions, while the cook had problems with the cauldron.
In the author’s subjective assessment, the Haiku scene looked the most developed and visually pleasing. The author also considered it the most stable of the three versions.
In the Luna version, the fisherman did not look convincingly engaged in fishing, some of the characters did almost nothing, and the guard tried to run away.
The author separately noted scaling problems: the buildings looked too small relative to the people, raising doubts about whether the characters would fit inside the castle. This observation applies to the scene shown, not to the model as a whole.
During the generation of the DeepSeek project, it stopped several times, and the author asked it to continue. The reason was not established. The author suggested a possible connection with OpenRouter or general service instability, but this remains an assumption and does not prove a property of the model itself.
In the DeepSeek village, the houses looked larger, the guards walked through the streets, the children ran in circles, and the cook’s animation seemed relatively convincing. The fisherman, by contrast, floated above the surface instead of fishing normally. The author also noted numerous visual errors and a lack of collisions.
Speed, cost, and resource usage
| Model | 3D mage | Fighting game | Village | Usage and cost |
|---|---|---|---|---|
| Haiku 5.5 | The author called the result the most developed | The main fighting game elements were present, but picking up items with the G key did not work or was unclear | The scene seemed the most stable and developed to the author | The evidence mentions a Claude subscription costing about 20 dollars |
| GPT-6 Luna | The staff, magical orb, and effects were present; problems with the hat remained | The project launched, but the author did not find clear attacks or working controls using JKL | The author noted unconvincing character behavior and scaling problems | The evidence mentions a Claude subscription costing about 20 dollars |
| DeepSeek V4.1 Flash | The result seemed primitive to the author | The author noted grabs, throws, dodges, and a boss | The scene contained character actions, visual errors, and a lack of collisions | A separate run cost about 22 cents; some requests reached 70 cents |
Directly comparing a subscription with per-request payments requires caution. The Claude subscription and the individual payments for DeepSeek belong to different access models, so the stated costs cannot be used as an exact estimate for the cost of the same amount of work.
Category results
- Best stability: Haiku, according to the subjective assessment of the author of this test.
- Best low-poly 3D mage: Haiku, thanks to the more complete development of the character and effects in the run shown.
- Best fighting game: DeepSeek, according to the author’s assessment, thanks to its dynamics, grabs, throws, dodges, and variety of scenes.
- Best village: Haiku, despite the lack of collisions and the problems with the cauldron.
- Fastest and cheapest run: this conclusion cannot be confirmed by the available evidence.
- Weakest game result: Luna in the fighting game shown, since the project launched but the author did not find clear controls or attacks.
If a prototype 3D scene with characters and basic animations were needed, the author would choose Haiku after this experiment. For a dynamic game prototype, the author would test DeepSeek, allowing time in advance to fix the mechanics and monitoring request costs. In the comparison shown, Luna was suitable for a visual sketch, but its game result required especially thorough testing.
What cannot be considered proven
The test does not confirm the universal superiority of one model over another. The outcome may have been affected by the wording of the prompt, the agent’s sequence of actions, the state of the project, the access interface, generation randomness, and other run conditions. Each task was tested in a limited number of variants, so the result cannot be considered an independent benchmark.
The assumption that roughly ten games could be mass-produced per day with a subscription costing about 20 dollars remains the author’s personal assessment. It does not account for review time, bug fixing, control testing, design, publication, support, or expenses for other tools.
The doubts about easy monetization of such projects should also be regarded as the author’s opinion. The author noted that the internet is already overflowing with similar prototypes, but this does not prove that a commercial result is impossible. Revenue depends on genre, quality, audience, promotion, and distribution method.
Who each model is suitable for
The author would choose Haiku for tasks where, in the comparison shown, the coherence of the result and subsequent manual refinement were important. The model did not eliminate all errors, but the author considered it the most consistent across all three scenarios.
The author associated DeepSeek with experiments in game logic because it was specifically in the fighting game that the author noticed grabs, throws, dodges, and more pronounced dynamics. In the author’s observations, the 3D mage was primitive, while village generation was accompanied by stoppages whose cause was not established.
The author would consider Luna for prototyping a visual idea, provided that the controls, animations, and object scaling were checked independently. In the comparison presented, the project launched, but the fighting game seemed unplayable to the author, and the village scene revealed problems with character behavior and proportions.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.
