AA-3D v1.0 Elo
Code
Artificial Analysis' 55-task evaluation of interactive three.js experiences created with Stirrup in offline sandboxes, with one attempt, up to 250 turns and three hours per task. Bradley-Terry ratings use equally weighted blind votes from a recruited human panel on eligible working scenes; GPT-6 Luna anchors the scale at 1000. Public votes do not enter the ratings.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 1199 |
| Claude Opus 5.5 | 1196 |
| GPT-6 Astra | 1164 |
| Gemini 4 Argon | 1153 |
| Claude Fable 5.1 | 1135 |
| GPT-6 Sol | 1121 |
| GPT-6.1 Sol | 1119 |
| Kimi K3.1 Pro | 1110 |
| Muse Spark 1.3 | 1033 |
| GPT-6 Luna | 1000 |
| Claude Haiku 5.5 | 998 |
| Grok 4.7 | 995 |
| MiMo-V2.6-Pro | 983 |
| Kimi K3 | 982 |
| Gemini 3.8 Flash | 966 |
| Qwen3.8 Max (0902) | 945 |
| GLM-5.3 Flash | 937 |
| Mistral Large 4 Preview | 870 |
| Step 5 Preview | 854 |
| GLM-5.3 | 783 |
| MiniMax M3 | 740 |
| Inkling | 583 |
| Nemotron 3 Ultra 550B A55B | 573 |