Claude Sonnet 5.5 Tops New AA-3D Scene Benchmark

Artificial Analysis has introduced AA-3D, a benchmark scoring 21 AI models on constructing interactive web scenes using three.js across 55 prompts. Based on over 10,000 blind human votes, Claude Sonnet 5.5 achieved the highest nominal rating at 1194 Elo. It sits in a statistical tie alongside Claude Opus 5.5 (1190 Elo), GPT-6 Astra (1165 Elo), and Claude Fable 5.1 (1131 Elo). Development costs ranged from $0.20 per scene for eighth-place GPT-6 Luna to $32.08 for Fable 5.1, with input tokens driving 69% of expenses. GPT-6 Sol proved cost-effective at $3.46 per scene with 1123 Elo, while Claude Sonnet 5.5 averaged $15 per run.

Open-weights alternatives lagged significantly, led by Kimi K3 (11th, 982 Elo) and MiMo-V2.6-Pro (12th, 977 Elo)—both over 200 Elo behind top proprietary models. Kimi K3 also spent 68 times as much as GPT-6 Luna for similar performance. Thinking Machines' Inkling and NVIDIA's Nemotron 3 Ultra scored under 580 Elo. Overall results correlated strongly with the Artificial Analysis Intelligence Index at a 0.92 Pearson coefficient (adding roughly 18 Elo per index point), outperforming comparisons to Terminal-Bench 4.0 at 0.81 and SciCode at 0.76. OpenAI entries beat index projections, with GPT-6 Luna, GPT-6 Sol, and GPT-6 Astra finishing 141, 93, and 46 Elo above expectations.

For practitioners using agent harnesses like the open-source Stirrup, visual iteration timing proved critical. Early screenshot inspection correlated strongly with score (0.82 Spearman). Top models inspected outputs after a median of 10 to 13 turns—reaching visual feedback by turn 20 in 89% of scenes—whereas 14th-to-20th ranked models waited 15 to 40 turns. MiniMax-M3 inspected scenes 54 times but ranked 19th because only 7% occurred by turn 20. Image capability also mattered: text-only GLM-5.3 scored 197 Elo below trend, whereas GLM-5.3-Flash scored 149 points above. Additionally, generating interactive controls correlated at 0.90 Pearson (top models built 15 to 20 controls versus 6 to 8 for bottom entries), while excessive file reading correlated negatively at -0.74. Turn counts (-0.27) and output tokens (0.33) showed minimal impact.