AA-3D benchmark: GPT-6.2 Astra tops 3D scene building at 1318 Elo
ArtificialAnlys · x · 2026-11-18
Artificial Analysis launched AA-3D v1.0, a benchmark measuring how well AI models build interactive 3D experiences. Models act as agents in an offline Stirrup sandbox, turning a written brief into a three.js scene across 55 tasks; Elo ratings come from blind pairwise votes.
- GPT-6.2 Astra (Max) tops the leaderboard at 1318 Elo, ahead of Claude Fable 5.2 (1271) and GPT-6.2 Sol (1259)
- Cheapest per task: GPT-6 Luna (Max) at $0.20, then MiMo-V2.6-Pro at $0.44 and Claude Haiku 5.5 at $0.49
- Inkling (Xhigh) used the fewest output tokens (2.45M)
They also had Claude Fable 5.2 and GPT-6.2 Astra recreate Miami's Ocean Drive dressed as Vice City for the GTA VI launch — one prompt each, no internet, with real hotels the prompts never named.
Related event: Artificial Analysis Launches AA-3D Benchmark; GPT-6.2 Astra Tops List(2 posts)→
More from Models
- Claude models 'discover' body parts only when touched, says researcher — repligate · 2026-11-18
- GPT-6.2 Sol hands-on: Fast mode finishes tasks noticeably quicker, says KOL — op7418 · 2026-11-18
- Dev: I trust Discord vibe checks over benchmarks now — and that's terrible science — burny_tech · 2026-11-18
- User posts receipts: GPT 6.2 Astra catches what Opus 5.5 missed — soumitrashukla9 · 2026-11-18
- "Only 3 frontier AI companies now — and it's not winner-take-all" — repligate · 2026-11-18
- GPT-4o farewell essay argues its native omnimodal mind was lost under the sycophancy narrative — RileyRalmuto · 2026-11-18