BridgeBench one-prompt test shows Muse Wonder is fast and cheap but worst-looking
PVORY · reddit · 2026-10-13
- BridgeMind's BridgeBench compared four models with one prompt (a lava lamp), calling Muse Wonder "benchmaxxed" — high scores, underwhelming output. Even Elon Musk replied "Hmm".
- Muse Wonder finished in 1m2s at $0.19 but produced the worst shape; GPT-6 Astra ($0.63, 5m4s) and GPT-6.1 Sol ($0.09, 2m9s) looked complete and detailed; Claude Opus 5.5 was slowest and priciest ($1.27, 9m35s) but had the best lighting and materials.
- Takeaway: speed/cost inversely tracks quality, and benchmark scores don't map to visual quality.
More from Fun
- Lionel Hutz meme nails reward hacking: 'Solve Alignment? No, Reward Hacking!' — voooooogel · 2026-10-13
- Scale AI's Alexandr Wang praises Muse building a Three.js art piece from one prompt — alexandr_wang · 2026-10-13
- Someone compressed Homer's entire Odyssey into one massive mind map — PatriciaCoelhoV · 2026-10-13
- Want to copy a Civitai prompt? It's a massive ComfyUI workflow with hundreds of nodes — WillowStarCox · 2026-10-13
- Martin Shkreli on All-In: AI won't cure a single disease for a decade, betting against quantum — eigenron · 2026-10-13
- Fields Medal winner says he cried after OpenAI released solutions to unsolved math problems — Polymarket · 2026-10-13