Cedric Chee Benchmarks Voxel Pagoda Scenes Across Sol, Fable, and Opus

On July 17, Cedric Chee published a batch of posts centered on the same voxel pagoda garden-style scene, comparing GPT-5.6 Sol with Fable 5 and, in a separate test, Sol ultra with Opus 4.6. The series is notable because it goes beyond output images: Cedric also described how he thinks these models should be used and pointed out unusual generation behavior he observed during testing.

Core tests

Cedric first ran a direct comparison between GPT-5.6 Sol and Fable 5 on a pagoda voxel garden scene, showing multiple settings including medium, high, and xhigh. He said he has been testing GPT-5.6 and Fable a lot recently and planned to publish more results the same day. In another post, he reposted a related voxel pagoda garden evaluation that also included Kimi-K3 (max), GPT-5.6 Sol, and Fable 5, but the provided material does not preserve the full ranking details.

Cedric’s observations

Cedric argued that GPT-5.6 Sol Pro feels less like a “builder” and more like an “architect,” meaning he sees its strengths in context engineering, discovery, planning, and system design. He also highlighted a surprising detail from a Fable 5 xhigh vs. high comparison: the model added three wing-flapping birds flying around the pagoda on its own, which he said was the first time he had seen that kind of spontaneous addition from the model.

Another comparison

Cedric also used a large voxel pagoda benchmark to compare Sol ultra and Opus 4.6. The clearest verifiable point from the available material is that Sol ultra completed the difficult scene in one shot within 30 minutes. Taken together, the posts show Cedric repeatedly stress-testing high-complexity scene generation while paying attention to output quality, unexpected details, and how each model fits different kinds of work.

2026-07-17 ~ 2026-07-17 · 6 related posts