Benchmark
Realism leaderboard
Board: Overall · Hands · Faces · Liquids · On-screen text · Multi-shot · Lip-sync · Camera · Animals · Style · Image to video · Animation
Look: Realism · Animation and 3D
Realism counts the 38 photographic prompts of the core battery and leaves out the anime and claymation tests.
Round: all · v1-vivix-2026-10 · v1-2026-10 · v1-2026-09 · anim1-2026-10
Clip shown: every model here on the same prompt, “camera-orbit-subject”, so the preview is a fair comparison, not a cherry-picked best clip.
| # | Model | Clip | Overall | 95% CI | Fidelity | Visual | Physics | Consistency | Audio | Attempts | Gen (s) | $/run | Runs |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |  Kling 3.0 Turbo
Kuaishou |
4.85 | ±0.12 | 4.71 | 4.88 | 4.79 | 4.84 | 5.00 | — | — | — | 38 | |
| 2 |  Seedance 2.0
ByteDance |
4.83 | ±0.15 | 4.78 | 4.89 | 4.59 | 5.00 | 5.00 | — | 266 | — | 37 | |
| 3 |  Google Gemini Omni
|
4.81 | ±0.12 | 4.80 | 4.86 | 4.79 | 4.80 | 4.82 | — | — | — | 38 | |
| 4 |  MiniMax H3
MiniMax |
4.78 | ±0.23 | 4.76 | 4.86 | 4.34 | 5.00 | 5.00 | — | 448 | — | 38 | |
| 5 | Seedance 2.5
ByteDance |
4.78 | ±0.16 | 4.68 | 4.88 | 4.69 | 4.77 | 5.00 | 2.00 | 233 | — | 37 | |
| 6 |  OpenAI Sora 2
discontinued OpenAI |
4.75 | ±0.18 | 4.72 | 4.83 | 4.25 | 5.00 | 5.00 | 1.00 | 217 | — | 38 | |
| 7 |  Grok Video
xAI |
4.71 | ±0.17 | 4.68 | 4.79 | 4.44 | 4.70 | 5.00 | 1.00 | 85 | — | 38 | |
| 8 | Google Veo 3.1
|
4.61 | ±0.27 | 4.57 | 4.72 | 4.25 | 4.79 | 4.38 | — | 122 | — | 38 | |
| 9 |  Runway Gen-4.5
Runway |
4.56 | ±0.16 | 4.46 | 4.64 | 4.49 | 4.64 | — | — | — | — | 38 | |
| 10 |  PixVerse v5.5
PixVerse |
4.51 | ±0.19 | 4.27 | 4.67 | 4.48 | 4.62 | — | — | — | — | 38 | |
| 11 | Kling 3.0
Kuaishou |
4.44 | ±0.2 | 4.28 | 4.66 | 3.97 | 4.80 | — | 1.00 | 190 | — | 38 | |
| 12 | MiniMax Hailuo 02
MiniMax |
3.85 | ±0.48 | 3.78 | 4.09 | 3.59 | 4.43 | 0.00 | — | 300 | — | 38 |
Attempts is the average number of generations needed before one was usable, the number that actually decides what a model costs you. 95% CI is computed from that model's own runs on this board, not assumed. On a category board with only 4 runs per model, a wide CI is the honest reflection of a small sample, not a bug. Two models whose CIs overlap are not meaningfully different at this sample size.
Cite this
Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.
Genativity. (2026). Realism leaderboard: AI video models ranked. Retrieved October 23, 2026, from https://genativity.com/leaderboard/realism<a href="https://genativity.com/leaderboard/realism">Realism leaderboard: AI video models ranked</a> (Genativity)Frequently asked
Which AI video model is best at Realism right now?
As of 23 October 2026, Kling 3.0 Turbo by Kuaishou ranks #1 on Realism on Genativity's benchmark, scoring 4.85/5 across 38 scored runs (±0.12 at 95% confidence).