The author asked GPT-6.2 Sol and Gemini 4 Carbon to create 3D content from the same open-ended prompt. This is a single, personal comparison and does not establish a general difference in performance or quality.
Tasks it suits: Exploring whether a model can turn a broad creative request into an expressive 3D work.
Tasks it does not suit: Model evaluations that require quantifiable conclusions, strict reproducibility, or production-grade quality assessment.
Applicable model versions: The original post labels the model GPT-6.2 Sol. The video was mistakenly labeled 6.1 Sol. The author later clarified that this was a typo and said the actual output came from 6.2 Sol. The post does not say whether this name refers to a generally available model or a preview route.
Applicable client, agent, or API: The author added that both models used the same harness, the Devin desktop app. This was not an API test.
Recommended reasoning level and parameters: The author said they used the highest reasoning level available for each model: xhigh for Sol and high for Carbon. Other parameters were not specified.
The author shared this prompt.
surprise me in 3d that makes me feel like you are agi modelRun each model in the same harness and submit the same prompt.
Use the highest reasoning level available for each model.
Compare the generated results. The original post shows different approaches in a video, but provides no scoring criteria or reproducible production steps.
The author said Sol took about 20 minutes, used about 6M tokens, and cost about $1.60. Carbon took about 40 minutes, used about 6.2M tokens, and cost about $2.90. The token counts and costs are the author's estimates. The post provides no billing statements or metering records.
The author only said the two models took different approaches and invited readers to judge for themselves. The specific outputs in the video are not described in the collected text, so this note does not describe details of the works or judge their relative quality.
This is a demonstration using one prompt, by one author, in one run. Time, token usage, and cost depend on the harness, reasoning level, and measurement method. They do not support conclusions about speed, price, or 3D capability rankings on other tasks.
To reproduce the setup described by the author, run the prompt above in the Devin desktop app, using Sol at xhigh and Carbon at high. The post does not specify the app version, environment, model routing, generated files, or measurement method. A reproduction may not match the video.
GPT-6.2 Sol