Video and Image Generation: Sora, Veo, Kling, Flux — 2026-09-08

Top developments

Kling AI Leads Text-to-Video Rankings in September 2026

As of September 2026, Kling v3 by Kuaishou leads the text-to-video leaderboard with an arena score of 1934, followed by Happy Horse 1.0 (1816) and Seedance 2.0 Fast (1747). This update reflects a growing preference for models that balance high-fidelity motion with cost-efficiency for commercial production. The shift underscores the rapid iteration in Chinese-developed models which are increasingly challenging Western incumbents in blind human vote benchmarks.

DOJ Argues Banning Copyrighted Training Data Would Hamper Progress

On September 4, 2026, the US Department of Justice stated that barring AI training on copyrighted material could "severely hamper" technological progress. In a filing, the DOJ argued that training models on copyrighted work falls under the same transformative use exception that protects journalism, such as that of The New York Times. This intervention comes amid ongoing litigation involving major AI labs and content creators, signaling a potential federal stance favoring broad training rights.

Local Newspapers Sue OpenAI and Microsoft Over Paywalled Content

In a move highlighting the ongoing tension between publishers and AI developers, local newspapers filed a lawsuit against OpenAI and Microsoft just two days ago. The suit alleges the tech giants used paywalled articles to train their models without permission or compensation. This case pushes back against administration claims that AI is a boon to local news, arguing instead that it undermines the financial viability of independent journalism.

Wan 3.0-Video Upgrades to 60B MoE Architecture

Alibaba’s Tongyi Wanxiang team released details on their August 2026 video generation series, with Wan 3.0-Video upgrading from a 14B dense model to a 60B Mixture-of-Experts (MoE) architecture. The model supports resolutions up to 1080P at 30fps and can generate videos up to 30 seconds long. It allows for complex multi-modal inputs, including up to 10 images, 5 video segments, and 5 audio segments as reference inputs, positioning it as a strong competitor in the open-weight space.

Local view

Chinese media outlets like Zhihu are actively comparing domestic tools such as Kling (Kuaishou), Jimeng (ByteDance), and Hailuo (MiniMax) against global rivals. Recent discussions highlight that while Jimeng has faced latency issues with its Seedance model, Kling and Midjourney are favored for specific creative tasks like traditional Chinese aesthetic generation. Users report that Kling offers superior control and cheaper iteration compared to Luma, making it preferred for volume production.

Context & numbers

• Arena Rankings (Sept 2026): Kling v3 leads T2V with an Elo of 1934; Happy Horse 1.0 follows at 1816.
• Pricing Benchmarks: Runway Gen-4.5 costs approximately $1.20 for a 10-second clip via API, while Kling 3.0 on third-party platforms costs roughly $0.75 for the same duration.
• Model Specs: Alibaba's Wan 3.0 supports 1080P resolution at 30fps with a max length of 30 seconds.

On the radar

• Andersen v. Stability AI Trial: The first US court ruling on whether AI training on copyrighted data constitutes fair use is beginning this week (September 8, 2026), setting a critical precedent for the industry.