GPT-6.2 Astra tops Blueprint-Bench 2, claims best physical 3D understanding among AI models
gdb · x · 2026-11-20
Greg Brockman relayed a claim from eval firm Andon Labs that GPT-6.2 Astra ranks #1 on Blueprint-Bench 2 — a benchmark where AI agents draw floorplans from photos of apartment interiors, testing physical 3D world understanding. GPT-6.2 Sol ranks #3. The claim comes via a quoted post, with no full eval report linked yet.
More from Models
- Claude models 82% off on RouterHub: Opus 5.5 at $0.72 in, Fable 5.2 at $1.43 in — zebird0 · 2026-11-20
- Fan-made chart claims MAI Code 2 matches 96% of Claude Opus 5.5 coding perf at 5% of the cost — marlene_zw · 2026-11-20
- Team claims Codex cracked 35 open Erdős problems in 2 months, some Lean-verified — IgorCarron · 2026-11-20
- Google's new AI reportedly cheaper and faster than Claude Opus 5.5 — HotMinuteUK · 2026-11-20
- MiniMax H3 video model reportedly blocks explicit prompts despite reference images — silleeni · 2026-11-20
- Google ends free Gemini Pro student subscriptions worldwide as model matures — Angaisb_ · 2026-11-20