What a 13-point Elo gap means: Wan 3.0 vs Seedance 2.5 win chance

A 13-point Elo gap means that, on the standard Elo scale, the higher model wins about 51.9% of head-to-head votes and loses 48.1%. That is Wan 3.0 (1,156) against Seedance 2.5 (1,143) on Artificial Analysis's text-to-video board (AA-Video-T2V v2.0). Out of 100 blind votes between the two, Wan would take about 52 and Seedance about 48.

The scale is an assumption, not a vendor statement: the board pages give Elo scores without a scale statement in the text read, so treat the percentages below as arithmetic on the usual Elo formula, where the expected score is 1 / (1 + 10^(-gap / 400)).

Gaps taken from the three boards

Each row is a pair of entries from the boards, with the gap computed from the scores as read. The win chance is the formula above, rounded to one decimal.

How to use the percentages

A gap under about 30 points is a small edge. At 27 points, the higher model wins 53.9% of votes, so about 46 in every 100 voters still prefer the lower one. A single clip from each will often disagree with the board. A gap of 100 or more starts to be a pattern: at 114 points the higher model wins about two votes in three.

The board also has error. A score rests on a finite number of votes, and a small gap between neighbours can flip. Do not treat rank 1 and rank 3 as a settled order when the scores are 13 points apart.

What follows for buying

If two models are within about 30 Elo of each other, let price, clip length and input limits decide, and run your own prompts for the final call. On Sume that means Wan 3.0 at $7.50 per minute at 720p against Seedance 2.5 at $34.67, for a 13-point gap. See the Wan and Seedance comparison for the full price table.

When the gap is large, as with Wan 3.0 against Kling 3.0 Omni on editing, the board is a better guide than a few test clips, though it still says nothing about your brand, your prompt style or your ratio. Run a small test to check the fit, in the way the leaderboard check post describes.

Gap checklist

Before you act on any gap, check three things. First, the board and the version: the text-to-video board is v2.0, image-to-video v1.0 and editing v1.1, and scores from different boards do not mix. Second, the date you read it, since a board moves as votes arrive. Third, whether the price column is for the size you will buy.

Then write down what a wrong pick costs you. If it is a 5 second clip you can regenerate for under a dollar, a 52% edge is not worth a long debate. If it is a 30 second 1080p Seedance 2.5 job, which costs $42.65 on Sume, run a cheaper test first.