The author used a short prompt to compare Gemini 4 Carbon with Fable 5.1 High on generating a 3D city of Oxford, showing the results side by side in a video. The author felt Carbon produced better architectural detail and lighting.
Suitable tasks: Generating an interactive 3D city of Oxford and delivering it as a single HTML file.
Unsuitable tasks: Projects that require a defined city area, building list, interaction behavior, performance metrics, or visual style. The original prompt does not specify these details.
Applicable model versions: Gemini 4 Carbon (as labeled in the post and video). The comparison model is Fable 5.1 High (video label).
Applicable client, agent, or API: Not specified. The video labels the model names but does not show a client or API.
Recommended reasoning level and parameters: Not specified. Fable 5.1 High is the comparison label in the video and does not indicate Carbon's reasoning level.
The original post uses the following single-line task as a model comparison prompt:
用 three.js 建一个可交互的牛津 3D 城市,单 HTML 文件English translation:
Build an interactive 3D city of Oxford with three.js, in a single HTML file.The original post shows the task prompt and a video of the generated results side by side. It does not explain how the prompt was submitted to each model or provide reproduction steps or source code.
The X page lists the author as Leo · Building w/ AI (@leo114119) and the publication date as October 30, 2026.
The author phrased the task as “Build an interactive 3D city of Oxford with three.js, in a single HTML file.”
The embedded video is about 8.5 seconds long and shows side-by-side labels “Fable 5.1 high” and “Gemini 4 Carbon.” The title “Oxford — The Dreaming Spires” is visible in the video.
The author says Carbon is clearly ahead in architectural detail and considers its rooftop architectural details and sunlight-driven lighting effects better, while also noting that the video is blurrier. This is the author's qualitative assessment of the video, not an independent evaluation result.
The post does not specify the client, API, Carbon reasoning level, generation time, whether the inputs were identical, number of iterations, or output file size.
This is one generation example and qualitative comparison posted by the author. It does not support conclusions about the models' overall relative performance in other 3D scenarios.
The short, compressed video and single viewpoint limit assessment of the models, interactive features, code quality, and runtime performance. The post does not show the HTML file or any interaction testing.
The prompt specifies only the technology, subject, interactivity, and single-file constraint. For reliable landmarks, navigation, camera controls, or performance requirements, specify those separately.
Collected on 2026-11-16. Opened the original X post through Tabbit and checked the author, page date, original prompt, model labels, and embedded video. The video was playable at the time of collection. It was not downloaded, no source code was obtained, and the models were not called for reproduction. The test environment and parameters were not disclosed.
Gemini 4 Carbon