To answer this, we mapped how both models think while they build Three.js games on DesignArena. We segmented their thinking traces into 800,000+ sentences using SaT (Frohmann et al.) and turned every sentence into an embedding using text-embedding-3-large (OpenAI).
Sentences that mean similar things land close together, so we reduced their embeddings to five dimensions using UMAP (McInnes et al.) and grouped them into topics using HDBSCAN (Campello et al.), choosing the clustering by its DBCV score (Moulavi et al.). Then we compared how much of each model's thinking falls in each one.