Priors adds Wojak-Bench 3D leaderboard
Priors has integrated Wojak-Bench 3D into its benchmark platform, tracking how frontier AI models convert meme drawings into printable 3D figures alongside subscription-adjusted task costs. Gemini 4 Carbon currently leads the community-voted Elo leaderboard with an 86% win rate, outpacing GPT-6 Astra and Claude Fable 5.2.
Physical-world output evaluation like 3D printing exposes model capability divergence that saturated text benchmarks fail to capture. Translating informal 2D sketches into watertight, printable geometry tests spatial reasoning under constraints where prompt adherence directly impacts real-world utility.
- –Gemini 4 Carbon leads decisively with a 1251 Bradley-Terry rating and an 86% win share across 1,740 pairwise comparisons.
- –GPT-6 Astra (1067 Elo) and Claude Fable 5.2 (1031 Elo) follow in the top tier, while models like Claude Sonnet 5.5 (682 Elo) and Grok 4.7 (636 Elo) drop off sharply.
- –Evals rely on human blind preference votes comparing rendered print files and physical prints across up to three revision rounds rather than synthetic LLM judges.
- –Priors maps these standings directly to subscription-plan cost calculations, giving developers visibility into performance per dollar on complex spatial generation tasks.
DISCOVERED
15m ago
2026-11-16
PUBLISHED
59m ago
2026-11-16
RELEVANCE
AUTHOR
AdamHoltererer