Seedance leads, Veo loses, Omni replaces — the video-generation arena resets
Seedance 2.0 holds #1 on Artificial Analysis Video Arena in both text-to-video and image-to-video. Veo 3.1 is below. Sora 2 is API-only after the consumer shutdown. Gemini Omni replaces the Veo branding with a 10-second cap. The video-generation competitive surface has been completely rearranged in five months.
Six months ago the video-generation pecking order was OpenAI Sora at the top, Runway and Pika as the credible alternatives, Google Veo as the well-funded but late entrant, and Chinese labs as a quality tier behind. Six months later: Sora 2 consumer app was shut down in March after burning $8-12M/month, Veo got renamed and replaced by Gemini Omni at I/O 2026, and Chinese labs lead the independent benchmarks.
Seedance 2.0 holds the Artificial Analysis Video Arena top spot in both text-to-video and image-to-video ahead of Kling 3.0, Veo 3.1, and Sora 2. Pairwise-comparison human evaluations. Same methodology as the text-model arena. The leaderboard is Chinese-lab-dominated.
How this happened in five months
Three structural shifts compounded. First: the Chinese video labs trained on chips they couldn't get through H100/H200 export controls, which meant they had to optimize architecturally rather than rely on raw scale. Architectural optimization translates to better $/quality at inference time, which is what the arena evaluates. Second: OpenAI's Sora 2 consumer-app shutdown removed the most-cited Western reference product from the market just as the Chinese labs were converging on quality. Third: Google's Gemini Omni Flash launched with a 10-second-per-clip cap that's a deployment-rationing decision, not a quality limitation — but the cap means the comparison-set head-to-heads aren't apples-to-apples.
The combination: the Chinese labs got better while the Western labs got constrained. Independent benchmarks reflect the result.
What the export-control framing now has to confront
The 2024-2025 export-control architecture was built on the premise that constraining advanced GPU exports to China would slow Chinese frontier-AI capability development. Video generation is the cleanest counter-example: the Chinese labs are leading the Western competitor set on an independent benchmark, while operating on a fraction of the H100/H200 fleet their Western counterparts deploy. Whatever they're using — Huawei Ascend 910C, Biren BR104, transshipped H-series via Singapore or Malaysia — the constraint did not work for this product category.
The 2026-2027 policy question is whether the response is tighter controls (the "close the loopholes" argument), a different framework (the "controls are counterproductive" argument), or an explicit shift to outcome-based metrics rather than input-based metrics. The current trajectory is ambiguous; the EU and US are not aligned, and the China response has been to keep shipping models that win benchmarks.
The pricing-model angle
Google's response to the compute supply constraint — clip-length caps on Omni Flash rather than full feature gating — is a different model than OpenAI's response to the same constraint. OpenAI shut down the consumer Sora 2 app to staunch the loss; Google is rationing capacity inside a still-live consumer product. Which approach generates better long-term consumer retention is unknown; the immediate competitive answer is that Google's Gemini app and YouTube Shorts have a working video-generation surface that Sora consumer no longer does. The economics will determine which becomes the dominant model.
TechCrunch — Google Gemini Omni turns images audio text into video → · 9to5Google — Gemini Omni create anything model starts today → · Medium / Analyst Uttam — Gemini Omni vs Seedance 2 →