Opus 5.5 tops BuildingBench at 86.8, but GPT-6 Astra matches it at 84% lower cost
ZhitingHu · x · 2026-09-25
EnactraAI updated its BuildingBench 3D reconstruction leaderboard: Opus 5.5 (max) jumped from Opus 5's 74.5 to 86.8 — the highest yet — ahead of GPT-6 Astra (ultra, 84.3) and Fable 5.1 (max, 81.4).
The cost gap is stark: GPT-6 Astra builds at $7.05 per building vs $45.47 for Opus 5.5 — just 2.5 points behind at 84% lower cost. GPT-6 Sol takes another tradeoff: ultra at 69.8 ($1.92/building) and max at 69.6 ($1.89), below GPT-5.6 Sol's 73.9 but roughly 75% cheaper.
A companion demo had Opus 5.5 and Astra independently reconstruct a square kilometre around Madison Square Park in Blender and Unreal from identical inputs, with no internet or human help — though Opus's pedestrians glide without moving their feet.
More from Models
- User claims Claude Opus 5.5 autonomously created a short film, no engines or references — RileyRalmuto · 2026-09-25
- Contrastive-LM releases CLM-v0.1-8B, a Qwen3-8B-based reranker trending on Hugging Face — Contrastive-LM · 2026-09-25
- Hesamation hints he's testing the rumored Claude Opus 5.5, calling it a return to a 'toxic relationship' — Hesamation · 2026-09-25
- Jev, a classification-not-generation model, goes viral with 1M-view intent demos — danshipper · 2026-09-25
- Qwen 3.8 27B Gets Stuck Writing 100+ 'Running Tests Now' Notes Without Doing Anything — Graemer71 · 2026-09-25
- Meta's Muse Spark 1.4 spotted on OpenCode's data page, launch may be near — kimmonismus · 2026-09-25