GPT 6.1 Soul, Sonnet 5.5, and Opus 5.5 in 3D scene creation: quality, time, and limit usage

A comparison of GPT 6.1 Soul, Sonnet 5.5, and Opus 5.5 shows how different AI agents handle environment setup, model imports, animation, and result verification. This is not a laboratory benchmark or a universal model ranking, but an analysis of three practical runs by the author involving different types of tasks.

All experiments used long prompts. They described the project, the sequence of stages, the required logic, and the final checklist. The agent received not only a request to “create a scene,” but also the criteria by which it was supposed to check its own work.

Methodology for comparing 3D scenes

The scenes differed in content, so model quality cannot be reduced to a single number. GPT 6.1 Soul assembled a magical scene with a character, staff, and effects. Sonnet 5.5 worked with an orc, an axe, and combat animations. Opus 5.5 received a simpler set of models, but had to configure fish movement, an aerator, particles, and water distortion.

The comparison considered four parameters: visual convincingness, correctness of object interactions, duration of autonomous operation, and consumption of available limits. The percentages and times refer to the author’s specific runs and do not indicate the models’ average performance across arbitrary projects.

GPT 6.1 Soul: mage, staff, and spell effects

GPT 6.1 Soul was run in XHigh mode. The agent received an AMD file with instructions for creating the project and imported models. According to the author’s observation, the task took 19 minutes 36 seconds and used 30% of the five-hour limit.

The final project included a starry background, a stone beneath the mage, audio files, and three types of magic. The mage held a staff and transitioned from a T-pose to another position, while the scene used lightning and fire effects. This required coordinating several independent components: skeletal animation, weapon positioning, lighting, sound, and visual effects.

The author called the result the best version of neural-network-generated 3D scenes they had seen before. This is a personal assessment, not a confirmed comparative metric. Artifacts remained in the scene, although in the author’s opinion some defects could be fixed by regenerating individual elements.

The character’s preparation also affected the result. The mage was created using a T-pose reference and a multi-view reference from all sides. The staff was processed separately so it could be controlled independently when raised and when effects were added. The character itself was created in a low-poly style. Among the shortcomings, the author noted weak lighting, four fingers instead of five, and an appearance they compared to a blue Santa Claus.

Sonnet 5.5: orc, axe, and combat animation checks

Sonnet 5.5 operated in XHigh mode on a low-poly orc in the aesthetic of 2000s games. The axe model was created separately first, after which the orc and weapon models were saved in GLB format and imported into the project to assemble the environment and animations.

The scene received breathing animation, sounds, grass, rocks, and low-poly trees. According to the author’s observation, the orc held the axe properly, but the weapon was too small relative to the character. The automatic assembly correctly linked the object to the hand, but did not preserve artistic proportions.

In one version, the orc hit himself. In another animation, he struck a rock with the axe, but the movement looked unnatural: the arm bent instead of making a straight swing. The author believes the result will become noticeably better after this animation is fixed. For now, this is an assessment of a specific scene, not proof that similar errors are consistently fixed after a rerun.

Sonnet worked for about an hour and a half. The author stopped the task after it reached 60% of the five-hour limit because the agent continued checking the axes through Playwright, even though, according to the author’s observation, the main work had already been completed. The run consumed 8% of the weekly limit.

Opus 5.5: aquarium and simple movement logic

Opus 5.5 was run in medium reasoning mode for an aquarium scene. The author chose this mode because the model was not required to create complex objects from scratch. The set included a castle, gravel, rocks, fish, snails, a crab, and plants. The main task was to configure movement and visual effects.

The work took 33 minutes. In this specific run, 32% of the five-hour limit and 5% of the weekly limit were consumed. These values are affected by prompt length, the number of tool calls, repeated checks, and environment settings, so they should not be extrapolated to every project with Opus 5.5.

The final scene included slowly swimming fish, an aerator with a connected hose and bubbles, water-induced image distortion, micro-particles, subtle plant movement, and snail and crab animations. In the author’s assessment, the fish movement resembled the behavior of live fish. The water distorted objects when viewed from above, while the bubbles created a noticeable wave effect.

Shortcomings remained. The particles were visible from above but barely appeared from the side. The plants moved slightly, but stood with their roots on the sand instead of being slightly embedded in it. The snail and crab moved; the author also suggested that the snail moves its eyes. The latter observation remains an assumption, not an established property of the scene.

What runtime and limit usage mean

The fastest run in this comparison belonged to GPT 6.1 Soul: it completed the work in 19 minutes 36 seconds. Opus 5.5 completed the task in 33 minutes. Sonnet 5.5 worked considerably longer, but its run included more verification iterations and was stopped by the user rather than ending at a natural point.

The percentages also require cautious interpretation. GPT 6.1 Soul used 30% of the five-hour limit during a short session, Opus 5.5 used 32% in 33 minutes, and Sonnet reached 60% by the time it was stopped. The percentage reflects not only duration, but also the intensity of the agent cycle, output volume, tool usage, screenshots, and repeated checks.

Why the agent continues working after the result is ready

The agent follows the specified criteria and available tools rather than the human feeling that “it is enough already.” When a prompt requires checking an object through Playwright, the model may continue the cycle until it finds a formal reason to end the check or until the user stops the process.

In the case of Sonnet 5.5, the author observed exactly this scenario: the main scene had been assembled, but the agent continued checking the axes. The cause could have been an overly broad validation requirement, repeated checking after every change, or the absence of a strict stopping condition. The precise internal cause cannot be established from a single run.

To control time, the prompt should include a separate completion block:

• first assemble the required objects and logic;
• then perform one checklist verification pass;
• fix only errors that affect the result;
• do not repeat checks on elements that have already been confirmed;
• finish after generating the final report.

The role of a detailed prompt and checklist

All three experiments used long prompts describing the project, stages, and verification requirements. This format is especially useful for 3D scenes, where the result depends on the order of operations. First, the agent needs to import the models, then place the objects, configure the connections, add animations, and only after that check the scene in the running environment.

A short request such as “create an aquarium with fish” does not establish quality criteria. A detailed instruction can separately describe swimming speed, aerator operation, bubble visibility from different angles, plant movement, and the behavior of secondary objects. The more requirements there are, the more important it is to divide them into mandatory and desirable ones; otherwise, the agent will spend the limit on secondary details.

Anthropic’s official documentation describes effort levels as a way to balance depth of work, time, and token usage. It also emphasizes that effort is not a strict budget, while a hard limit is set by separate environment parameters. In this comparison, this approach helps explain why medium reasoning mode was sufficient for Opus in the aquarium’s logic, while XHigh was used for longer tasks involving characters and animation.

Which model to choose for an 3D task

In the author’s assessment, GPT 6.1 Soul performed best in a task requiring a cohesive demonstration scene with a character, weapon, sound, and effects. This option is suitable for quickly prototyping a game ability or presentation scene if the developer is prepared to fix artifacts manually.

Sonnet 5.5 proved useful for an intensive agent process involving the import of multiple models, an environment, and checks through a browser tool. Its run showed that the model can work on a single scene for a long time, but requires control over the task boundaries. Without an explicit completion condition, unnecessary checks can significantly increase limit usage.

In this experiment, Opus 5.5 was used for a scene whose complexity lay not in modeling, but in object behavior and effects. The aquarium showed that a convincing scene is composed of small components: slow fish movement, bubbles, water distortion, micro-particles, and secondary animations.

• GPT 6.1 Soul is worth considering for cohesive scenes with a character and visual effects.
• Sonnet 5.5 is suitable for long agent tasks requiring asset imports, environment assembly, and tool-based verification.
• Opus 5.5 is reasonable to use for complex behavior logic and scene atmosphere when the basic models are already prepared.

What to refine manually after generation

None of the described runs produced a result that could immediately be considered a final game asset without verification. After generation, it is necessary to check object scale, weapon attachment points, hand positions, attack trajectories, collisions, lighting, and the appearance of effects from different angles.

For the mage scene, the fingers, lighting, and character artifacts require particular attention. In the orc scene, the axe should be enlarged and the strike against the rock reworked so that the arm follows a natural trajectory. In the aquarium, plant placement, side visibility of particles, and whether the fish movement matches the chosen game style should be checked.

The main conclusion of the comparison concerns not any individual model, but the combination of assets, a detailed prompt, staging, and manual acceptance. An AI agent can already assemble an expressive 3D scene and connect it to animation, but the quality of the final result is still determined by the precision of the task description and the care taken during verification after the work is complete.