AI video leaderboards disagree on the No. 1 model

Two leaderboards that rank AI video models from blind human votes named different winners when they were read on November 12, 2026. Artificial Analysis's text-to-video board put Alibaba's Wan 3.0 first at 1,175 Elo. Arena's text-to-video board put Google's Gemini Omni 1.2 Flash first at 1,540, marked preliminary. Both readings were correct. That is the whole lesson: a leaderboard is a snapshot of who won a set of blind comparisons on a particular day, not a fixed verdict on which model is best.

This is a practical guide to reading those boards, to understanding why they disagree, and to using them for what they can actually do: build a shortlist before you spend money.

The two boards named different winners today

Artificial Analysis's AA-Video-T2V v2.1 board, read on November 12, 2026, ranks video models on generating video with audio from a text prompt. It put Wan 3.0 first at 1,175 Elo, ahead of Utopai X at 1,170, MiniMax H3 at 1,161, Dreamina Seedance 2.5 at 1,155 and Gemini Omni Flash 1.2 at 1,154. The board lists 34 models and says seven were added in the last 30 days.

Arena's text-to-video board, read the same day and showing a board date of November 7, 2026 with 891,840 votes across 50 models, put Gemini Omni 1.2 Flash first at 1,540, marked preliminary, ahead of Gemini Omni 1.1 Flash at 1,513 and an unversioned gemini-omni-flash entry at 1,509. Wan 3.0 sits eighth on Arena at 1,476. The model Arena puts first is fifth on Artificial Analysis at 1,154: the two boards disagree about the leader and about the same model's place in the field.

A third board names a third leader. Artificial Analysis's image-to-video board, read on November 12, 2026, put ByteDance's Dreamina Seedance 2.5 first at 1,161. On that board Wan 3.0 is 11th and Gemini Omni Flash 1.2 is 14th.

How a blind-vote Elo board is built

Both boards work the same way in outline. Two models generate a video from the same prompt. A voter sees both clips without knowing which model made which, and picks the one they prefer. Each vote nudges both models' ratings: beating a higher-rated model earns more than beating a lower one.

Artificial Analysis describes its rankings as "an Elo rating system derived from blind comparisons judged by our validated private evaluators." The underlying math is not the 1960s chess formula. According to Infer's guide to how these boards work, it is Bradley-Terry maximum likelihood estimation, rescaled afterward so the numbers look like Elo. One vote counts for almost nothing on its own; Artificial Analysis shows the sample behind each score, and Wan 3.0's first place on the text-to-video board rests on 7,651 votes. The board recomputes as votes arrive, so a number you read is at best hours old, which is why every figure in this article carries the board name and the date it was read.

Why the two boards disagree

Four things separate the two boards, and none of them is a mistake.

The vote pools differ. Artificial Analysis uses a private panel it calls validated evaluators. Arena is open: it showed 891,840 votes across its text-to-video board. Different people vote, with different preferences.

The rosters and cutoff dates differ. Arena's board carried a date of November 7, 2026. Artificial Analysis's board says seven models joined in the last 30 days. A model that arrived after one board's last update cannot be ranked on it, and the two boards update on their own schedules, which is why Gemini Omni 1.2 Flash is first on Arena at 1,540 and fifth on Artificial Analysis at 1,154.

The toggles differ. Artificial Analysis splits audio from silence into separate boards. Switch its text-to-video board to the silent version and the winner changes: HeyGen Video 1 leads at 1,146, and Wan 3.0 falls to fourth at 1,136. A single switch on one board produces a different No. 1 than that board's own default.

The scales differ, and they get re-based. Artificial Analysis anchors its scale at Kling 3.0 1080p Pro fixed at 1000; Arena's scale is its own. The version number on the board, v2.1, is the tell that the scale has been reset: an Elo from an earlier version is not comparable with a current one, and an Arena score is not comparable with an Artificial Analysis score. Infer documented the drift on a single model: Kling 3.0 1080p Pro sat at 1,248 Elo and third place on a cached Artificial Analysis snapshot dated April 30, 2026, and at 1,111 Elo and sixth place on the live board Infer observed on July 22, 2026, an 83-day gap with a 137-point swing. Nothing about the model changed. The field around it did, and Elo is relative, so an unchanged model drifts as new entrants take points from everyone they beat.

What Elo does not measure

A high Elo means voters preferred a model's clips more often in blind comparisons. It does not mean the model is right for your job.

It does not measure whether your product label survives a shot, or whether a character looks the same from cut to cut. It does not measure whether a finished piece of work is any good; it measures which of two clips a voter preferred, one prompt at a time. It says nothing about native audio, clip length, resolution, price or the control features you need.

The gaps between neighboring ranks are smaller than the numbers suggest. Infer's conversion table puts a 20-point Elo gap at about a 53% expected win rate for the higher-rated model, a 40-point gap at about 56%, and a 137-point gap at about 69%. Two models a few points apart are, on any single prompt, close to a coin flip. Artificial Analysis shows each rank as a range for the same reason, and Runway's developer platform, checking the boards on November 3, 2026, noted that MiniMax H3 sat anywhere from second to sixth on the text-to-video board and Seedance 2.5 from third to seventh.

How to use a board to build a shortlist

Read a board for what it can do: narrow the field.

Pick the board that matches your input. A text-to-video ranking does not carry over to image-to-video, because Artificial Analysis scores each modality as its own arena, and the leaders differ. Read the rank range, not the single position. Then check price, because equal-quality neighbors can cost very different amounts: Runway noted that two adjacent models on the text-to-video board cost $4.80 and $34.12 per minute of video on their creators' own APIs.

Vidwip's leaderboard roundup, updated on November 12, 2026, does the thing the boards themselves skip: it puts Arena's and Artificial Analysis's scores side by side and adds columns for clip length, native audio, resolution and price per second. Read it knowing that Vidwip is a vendor, and lists its own tool first and unranked because it is not a model. Its own conclusion is the point: no single model wins every board, so use the ranking to build a shortlist, not to pick a winner.

The checklist to apply instead

Once a board has narrowed your list, judge the models on what the boards leave out.

Clip length: Seedance 2.5 and Wan 3.0 both reach 30 seconds, while Gemini Omni 1.2 Flash runs 3 to 15 seconds. Native audio: most of the leaders generate it, but confirm it on the model you pick. Resolution: MiniMax H3 reaches 2K, and Veo 3.1 reaches 4K on Google Cloud, while several leaders top out at 720p. Price per second: Gemini Omni 1.2 Flash is the cheapest of the leaders at $0.07 a second on Google's own API at 720p, with MiniMax H3 at $0.08 a second at 768p. Control features: references, keyframes and prompt-based editing differ more between models than their Elo scores do.

The boards answer one narrow question. A production choice needs several, and the checklist is where the rest of the answer lives.

Both boards are free to read at artificialanalysis.ai/video/leaderboard/text-to-video and arena.ai/leaderboard/text-to-video, and Arena lets you vote in the matchups yourself.