Blog›Benchmarks

AI Video Models, September 2026: Arena Elo vs Human Ratings vs Real Usage

Gemini Omni leads the arena, Seedance 2.0 wins per-clip human ratings, and Veo 3 dominates usage. Three datasets, three different winners - and how to choose.

AI video generation is moving faster than any other modality right now, and the picture you get depends on where you look. The community arena crowns one family, human raters scoring individual clips prefer another, and prompt writers vote with their keyboards for a third. Here is the September 2026 state of AI video, grounded in two datasets we publish daily: the text-to-video leaderboard and Wikiprompt's own corpus of 6,900+ video prompts with rated results.

What the arena says

On today's benchmark snapshot, the top of the text-to-video arena by community Elo:

Gemini Omni 1.1 Flash (1515) and Gemini Omni Flash (1512) - Google's omni-modal line took the video crown this summer.

Flux 3 Video (1495) - Black Forest Labs' expansion from images into motion, the strongest open-adjacent entry.

Seedance 2.0 / 2.5 (1479 / 1476) - ByteDance's pair keeps the head-to-head race tight.

MiniMax H3 (1460) - the Hailuo lineage, consistently underrated in western coverage.

What human raters say, clip by clip

Our corpus scores each prompt's actual output on a 1-5 "wow" scale. Average wow across the video models with 350+ rated prompts:

| Model | Prompts | Avg wow |

|---|---|---|

| Seedance 2.0 | 1,876 | 3.47 |

| Kling | 614 | 3.38 |

| Wan 2.1 | 423 | 3.26 |

| Sora | 369 | 3.14 |

| Hailuo | 937 | 3.09 |

| Seedance 2.5 | 827 | 2.94 |

| Veo 3 | 1,409 | 2.84 |

Two things jump out. Seedance 2.0 outscores its own successor in per-clip wow - a reminder that newer is not always more impressive on the prompts people actually write. And Veo 3 ranks lower here than on arenas: its strength (native audio, dialogue) is invisible in silent-clip ratings, which is exactly why you should read multiple boards before picking a tool.

What prompt writers actually use

By volume, Seedance 2.0 (1,876 prompts) and Veo 3 (1,409) dominate the corpus; Wan 2.1 owns the local/open niche and Veo 2 remains the budget workhorse via the Gemini API. Kling punches above its volume in quality (3.38 wow on 614 prompts).

How to choose in practice

Cinematic realism with sound: Veo 3 - write the audio into the prompt.

Best average clip quality per prompt: Seedance 2.0 - see the Seedance guide.

Open weights / local: Wan 2.1, with 2.2 improving motion.

Character motion and choreography: Kling.

Watch next: Gemini Omni's arena lead and Flux 3 Video - both are new enough that their prompt culture is still forming, and our daily snapshots will show whether the Elo holds.

Every number above updates daily: text-to-video leaderboard · benchmark history · browse the clips behind the ratings in the Prompt Arena or per model at Seedance, Veo 3, Kling, Wan 2.1.