The author says Hy4 (max) completed the task in about 6 minutes at an estimated cost of about $0.03, while GPT 6.1 Terra (max) took about 16 minutes and cost about $0.44. The author considers Hy4 faster and cheaper, but says its 3D coding quality still trails Sonnet 5.5 and DeepSeek Flash 1017. Both time and cost are the author's estimates; the post does not provide the full prompt or a verifiable measurement method.
Tasks this comparison can inform: Generating a 3D medieval castle in a single HTML file; understanding the author's time, estimated cost, and subjective impressions for this particular task.
Tasks it cannot be generalized to: General web development, other 3D scenes, reliability or ability to make multiple rounds of revisions, and overall coding ability.
Model versions covered: Tencent Hy4 (max); GPT 6.1 Terra (max).
Test environment or client: The author provided a Tencent AI Studio Web Code share page and a Terra result page. In a reply, the author said this run was done by a bot and recommended using Hy4 with Web Studio rather than the API. It is not possible to confirm whether the two execution environments were equivalent.
Reasoning levels and parameters: Both are labeled max; temperature, random seed, context, code execution environment, and other parameters are not specified.
The author says both models used the same prompt to build a 3D medieval castle in one HTML file, and posted a side-by-side demo video. The original post says “full prompt below,” but the visible expanded replies do not include the full prompt. The post also does not specify when timing began or ended, how costs were calculated, or how many runs were performed.
Figures reported by the author: Hy4 took about 6 minutes and cost about $0.03 (explicitly marked as an estimate); Terra took about 16 minutes and cost about $0.44. Based on these approximate figures, Hy4 took about 38% as long as Terra and cost about 7% as much.
Observable outputs: The original post's video thumbnail shows the two castles side by side. The Hy4 scene on the left has several densely arranged towers with red pointed roofs and trees around them; the overall scene is darker. The Terra scene on the right shows a small island surrounded by water, with stone walls and red pointed roofs; the overall scene is brighter. The post also includes demo videos for each model, but during collection only the thumbnail images were checked; the full interactions and resulting code were not reviewed frame by frame.
Author's assessment: The author says Hy4's 3D coding performance has not yet reached Sonnet 5.5 or DeepSeek Flash 1017, but it is fast and inexpensive; in a reply, the author also said GPT was better in some respects. These are the author's impressions, not conclusions measured using a standardized rubric.
Feedback in replies: Some users said their Hy4 requests had run for nearly 30 minutes or 2 hours without finishing. The author speculated that launch traffic was responsible and said their own bot completed the task in under 10 minutes in the morning. This discussion suggests that wait times may depend on load at the time and the usage path, so a single observation should not be treated as a stable latency figure.
| Item | Hy4 | GPT 6.1 Terra |
|---|---|---|
| Reasoning level | max | max |
| Prompt summary | Build a 3D medieval castle in a single HTML file | Same, according to the author, using the same prompt |
| Time | About 6 minutes | About 16 minutes |
| Cost | About $0.03, marked by the author as an estimate | About $0.44 |
| Output link | Tencent AI Studio Web Code share page | Terra castle demo page |
The only conclusion supported by this record is that the author reported Hy4 as faster and less expensive in this single-task comparison, while preferring some aspects of Terra's output. The original post includes side-by-side images of the results and output links, but lacks the full prompt, repeated tests, a standardized scoring method, and a cost breakdown. Runtime conditions and load were also uncontrolled. The results therefore cannot establish a general performance gap or a stable service speed.
The visible prompt summary in the original post is “build a 3D medieval castle in a single HTML file.” The links to both outputs are included in the raw data table. Reproducing the reported speed and cost would require obtaining the full prompt, timing rules, cost basis, and execution environment, none of which the author disclosed, then running multiple trials under the same conditions. The available material is insufficient for an exact replication.
Hy4