Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresKaggle and HeyGen's arena for HeyGen's Code2Video design briefs: each model writes an animated video composition for a brief, and people vote blind between two models' rendered videos (A, B, both good or both bad). Kaggle reports an Elo rating from those votes, recomputed about every ten minutes, with an interval whose level it doesn't state. Unlike Code2Video-AA, which uses a Gemini judge, the ratings come from human votes.
41st strongest of 803 fitted benchmarks
Signal-to-noise +19.74
213th hardest of 803 fitted benchmarks
Capability needed 163.7
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon | high | 1599 | |
| 2 | Claude Opus 5.5 | high | 1589 | |
| 3 | GPT-6 Astra | high | 1561 | |
| 4 | Gemini 4 Argon | high | 1550 | |
| 5 | Claude Sonnet 5.5 | high | 1538 | |
| 6 | Claude Fable 5.1 | high | 1524 | |
| 7 | GPT-6.1 Sol | high | 1514 | |
| 8 | Claude Haiku 5.5 | high | 1474 | |
| 9 | GLM-5.3 | high | 1449 | |
| 10 | Kimi K3 | high | 1434 | |
| 11 | Gemini 3.8 Flash | high | 1415 | |
| 12 | GLM-5.3 Flash | high | 1381 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).