Alibaba released Wan 3.0 in August 2026. On the Artificial Analysis text-to-video arena, read 21 September 2026, it scores 1229 Elo against 1088 for Google's Veo 3.1 — and it bills half as much per minute of output.
That is an unusual shape for this market. The cheaper model is normally the worse one.
The two side by side
Elo here is blind pairwise human preference, with the confidence interval the board publishes. It is not a capability benchmark, and it is not our own rating recipe — it is an ordering anchor, and we treat it as one.
Where Veo still wins
Three places, and they are not small.
4K. Wan 3.0 tops out at 1080p. If you need 4K masters, Veo 3.1 is the only one of the two that produces them.
Latency. This is the real trade. Wan 3.0's median end-to-end request on its served endpoint is 239 seconds — close to four minutes — against 42.6 seconds for Veo 3.1. Wan is 5.6× slower to return. For batch rendering that is irrelevant. For anything interactive it is disqualifying.
Native audio. Veo 3.1 generates synchronised audio with the video.
Where Wan wins
Clip length, and it is not close: 2 to 30 seconds in a single generation against Veo's fixed 4, 6 or 8. Veo chains scene extensions past that, but each extension is another generation and another charge. A 30-second Wan clip is one request at $6.00. Thirty seconds of Veo is four generations at roughly $12.00.
Wan also runs at 30 fps against Veo's 24.
What this actually means
If you are rendering short social clips at 1080p and can wait, Wan 3.0 is both better-rated and half the price. If you need 4K, synchronised audio, or a response inside a minute, Veo 3.1 earns its premium.
One model beats both on price: MiniMax H3 sits at 1220 Elo — nine points behind Wan — at $7.80 per minute, and publishes open weights for its 768P tier under the MiniMax H3 Community Licence. It is the only open-weight model in the arena's video top five.
Prices and latency are read from each model's served endpoint; Elo, confidence intervals and vote counts from the Artificial Analysis video arena on 21 September 2026. Every figure carries its source and date on the model pages.