GPT-6.1 Terra in practice: a test of 3D scene, game, and website generation

According to the author of the practical test, GPT-6.1 Terra works with text, images, files, and computer tasks. The author also indicated a context of up to 1 100 000 tokens and no support for video as input data. The official description linked in the test materials associates the model with visual analysis, computer control, and programming. The results below relate to a single practical session and are not a universal ranking.

The author's main conclusion is that GPT-6.1 Terra quickly assembles working prototypes of websites, games, and interactive scenes.

The details, animation, and object behavior still have to be checked manually.

What is known about GPT-6.1 Terra

The test materials state that GPT-6.1 Terra is available in ChatGPT Work and Codex. They also provide the model identifier for the API: gpt-6.1-terra. The stated standard API cost is 1 dollars per 1 million input tokens, 0,05 dollars per 1 million cached input tokens, and 5 dollars per 1 million output tokens. These figures apply to the API pricing model, not to a subscription with usage limits.

The test author indicated one more important parameter: a context of up to 1 100 000 tokens. Such a capacity, if the specific interface and plan actually provide the stated limit, makes it possible to keep a large project, source code, images, and technical requirements in a single work session. Video was not used in the test scenario, so the model was evaluated on text instructions, files, and the visual output of projects.

At the time of the evaluation, the author did not have enough independent benchmarks to assess the model. The manufacturer's official metrics describe standardized developer tests, while the user experiment shows the model's behavior in several specific tasks. These two types of data should not be mixed.

Practical test conditions

The author conducted the evaluation under a subscription costing 20 dollars and enabled XH reasoning mode. The program included four projects: an 3D scene with an orc, an aquarium, a Mario-style game, and a website design with an interactive 3D object.

The experiment measured creation time and subscription limit usage. These figures relate to a specific session and mode settings, so they should not be considered standard for every user: the result depends on the length of the request, the number of iterations, the project size, and the actions the model performs in the work environment.

Across all tasks, the author reported using about 14% of the five-hour limit. He also noted a slow decrease in the available resource and considered the subscription convenient for a series of prototypes.

3D orc: detail and animation

The model created the orc scene in 3 minutes 14 seconds. According to the author's observation, the character's silhouette, ears, tusks, and hair were distinguishable in the result. For a quick prototype, this set of features is sufficient: the viewer understands who is in the scene, even if the model does not withstand detailed artistic scrutiny.

Problems appeared in individual elements. The nose looked more like a dog's than an orc's, and the axe came out blunt. Such defects do not interfere with demonstrating an interactive scene, but they reduce the quality of the result when the project requires a recognizable character design or game-ready assets.

In the test, the axe-swing animation worked adequately: the orc moved from top to bottom, and the main action was clear without additional explanation. The body language looked weaker. In the author's subjective assessment, the limbs barely bent, making the character move like a tin soldier. This observation applies to the specific project, not to an official assessment of the model's capabilities.

The author called the result acceptable for a younger, non-frontier model. The scene's practical value lies elsewhere: according to the reviewer's assessment, GPT-6.1 Terra quickly assembles the foundation of an 3D prototype, but a designer or developer still has to check the anatomy, joint points, movement balance, and consistency with the reference separately.

Aquarium: objects, effects, and movement physics

The aquarium took 3 minutes 51 second to create. According to the author's observation, the rocks and plants looked acceptable, while the fish had enough detail for a demonstration scene. The space included a two-tower castle, a bubbler, waves, water, and large, slowly moving snails.

The fish mostly looked like fish and moved plausibly: they slowly waved their fins and maintained a recognizable trajectory. The author also noticed distorted waves, small particles, and strange circles on some objects, which he interpreted as something resembling mouths.

The crab showed a more noticeable behavioral error. It moved backward, although in the author's opinion this character was expected to move sideways, and its legs formed incorrectly. The scene demonstrates a limitation revealed in this test: the model creates many objects and effects simultaneously but does not always correctly follow the biomechanics of each creature.

According to the author, the aquarium turned out well, with no major errors in the tested scenario. For a stricter evaluation, he suggested making the task more complex by adding a school of fish, more types of movement, and additional visual effects. This approach would make it possible to test the model's capability boundary, since individual objects and simple trajectories do not fully characterize it.

Mario-style game

The game was the longest project. Creation took 9 minutes 33 seconds and used 7% of the five-hour limit and 1% of the weekly limit. According to the author's description, the finished prototype included character movement, crouching, running, enemy animation, secret areas, a mushroom, points, transformation into a white character, fireballs, and a finish with a flag and castle.

During testing, the fireballs destroyed enemies, the falling-into-a-pit animation triggered, and the flag moved slightly and reacted to the wind. The author noticed no obvious bugs in the tested scenario. The game did not undergo full testing: not all action combinations, extended game sessions, different screen resolutions, or recovery after errors were checked.

The graphics looked unusual, but the character was low-poly, and the level of detail remained low. This result is suitable for a prototype or mechanics demonstration. A commercial release would require separate work on visual style, collisions, level balance, sound, accessibility, and progress saving.

Website design and an interactive 3D object

In the design test, the model quickly created a website and used about 2% of the five-hour limit and about 1% of the weekly limit. The author rated the fonts, visual presentation, and responsive version positively. In the mobile version, the 3D object displayed better, and he found the combination of lime green and pink colors successful.

The main technical shortcoming appeared when rotating the 3D object: part of the geometry was clipped. The container, canvas dimensions, camera, scene depth, block overflow, and the object's behavior on different screens should all be checked.

The visual concept seemed typically AI-generated to the author. In his opinion, with a precise prompt the model can handle design tasks reasonably well, but without brand guidelines, references, and several rounds of selection, the result lacks individuality.

Another shortcoming concerned interactivity. In the version tested by the author, practically every website element opened an email client to send a message. Such a template creates the appearance of functionality but does not replace real forms, request processing, CRM integration, or analytics event verification. This is acceptable for a demonstration landing page, but a working website would require full-fledged scenarios instead of such links.

How much resources the four projects required

The test showed that the tasks differed not only in prototype quality but also in creation time. The 3D orc and aquarium each took less than four minutes to create. The game required noticeably more time because of the number of mechanics and states. The website came together quickly, but speed does not eliminate the need to check responsiveness, navigation, and the actual actions of buttons.

• The 3D orc took 3 minutes 14 seconds and 2% of the five-hour limit.
• The aquarium took 3 minutes 51 second and 3% of the five-hour limit.
• The game required 9 minutes 33 seconds and 7% of the five-hour limit.
• The website used about 2% of the five-hour limit.
• Total usage in the experiment came to about 14% of the five-hour limit.

Who GPT-6.1 Terra is suitable for

According to the test author's assessment, GPT-6.1 Terra is suitable for frontend and backend tasks, rapid vibe coding, creating interactive demos, and testing ideas before full-scale development. The developer does not get a finished product requiring no oversight, but rather a starting version that can be run, viewed, and gradually refined.

Within the described experiment, the model proved useful where the speed of the first result matters: when creating a website prototype, game mechanics, a simple 3D scene, a visual concept, or an educational example. Reliability is lower in tasks where precise anatomy, complex physics, detailed graphics, stable behavior of many objects, and error-free business logic are critical.

Subscription or API

The API is more convenient for products, automation, and calculating costs based on the actual number of tokens. A subscription is suitable for interactive work in which the user repeatedly changes requirements, asks to check the result, and runs different projects in one interface.

The author considers the 20-dollar subscription more cost-effective than paying for model usage through the API specifically for this type of experiment. This is his assessment of a particular session, not a universal rule. With a small number of requests, the API may be more economical, while with many iterations, a subscription may provide a more predictable budget if the plan's limits cover the required amount of work.

Final assessment of GPT-6.1 Terra

The practical test demonstrates what the review author sees as GPT-6.1 Terra's strength: the model quickly turns a description into a working interactive prototype. In his session, it assembled an 3D orc, an aquarium, a game, and a website while preserving basic functionality and staying within a small amount of limit usage.

The limitations are noticeable in the details. The orc bent its limbs poorly, the crab's characteristic movement was incorrect, artifacts appeared on the fish, the website's 3D object was clipped, and the interactive elements often boiled down to launching an email client. The result can be used as a foundation, but not as a final production version.

Frequently asked questions

How long did GPT-6.1 Terra take to create the projects?

According to the practical test, the 3D orc took 3 minutes 14 seconds to create, the aquarium took 3 minutes 51 second, and the Mario-style game took 9 minutes 33 seconds. The author described the website as created quickly, without giving a separate exact time.

Does the model support video?

In the test description, the author indicated support for text, images, and files, but not video as input data. The official description linked in the test materials associates the model with video analysis and editing tasks in a computer environment, but this is not equivalent to a guarantee that a video file can be uploaded in every interface.

Can the created game be considered finished?

No. The main mechanics worked in the experiment, but the author tested a limited scenario. Before publication, collisions, performance, control on mobile devices, state saving, and all game transitions would need to be checked.

Is it worth choosing a subscription instead of the API?

For a series of interactive experiments, the author chose a subscription and considered it more cost-effective. For an automated product, the decision should be made after calculating token usage, the number of requests, limits, and integration requirements.