The mid-2026 shift: OpenAI exits consumer video, Google's Omni rebrand, and the Chinese surge

The video-model story since version 11 is one of realignment at the top, and the headline is who left. This section records the landscape as of late July 2026.

OpenAI exited consumer video. The Sora consumer app went dark on April 26, 2026, and OpenAI confirmed the Sora API will be discontinued on September 24, 2026. The reported causes are brutal and instructive: operating costs around a million dollars a day against a small fraction of that in revenue, downloads that fell from a 3.3-million peak to roughly a million, a collapsed Disney licensing deal, and unresolved copyright and deepfakeDeepfakeA synthetic image, audio, or video of a real person or event made to look genuine, a category the EU AI Act now requires be disclosed. problems. Sora continues only as an internal 'world modelWorld modelA model that learns an internal simulation of how a scene evolves, aiming to keep physics, objects, and space consistent over time. A frontier for long, coherent video.' research effort. The first major lab to ship a flagship generative-video product became the first to retire one.

Google rebranded rather than iterated. At Google I/O in May, Google did not ship a 'Veo 4.' It announced Gemini Omni, an any-to-any multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. family with conversational, multi-turn editing and physics-aware generation, and shipped the first model, Gemini Omni Flash, the same day to the Gemini app, Google Flow, and free inside YouTube Shorts. As of July, Gemini Omni Flash sat at number one across the major Artificial Analysis video arenas for text-to-videoText-to-videoGenerating video from text alone, with no starting image. and image-to-videoImage-to-videoStarting from a still image and generating a video that animates it according to a prompt..

The Chinese labs surged. Kuaishou's Kling 3.0 Turbo, on June 17, added fast previews, multi-shot prompting, and native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. with lip-sync in five languages, and the company raised a round reported to value it around 18 billion dollars, with Alibaba and Tencent among the backers. ByteDance's Seedance 2.5, announced June 23 with APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access in mid-July, generates native single-pass 30-second clips with no stitching, accepts up to 50 reference inputs, and outputs 4K. And an Alibaba unit's model, released under the name HappyHorse, climbed to the top of the open video arenas by generating audio and video jointly in a single forward pass.

The reliability caveat, restated. This corner of the field is now thick with search-optimized content asserting releases that never happened: a 'Veo 3.2,' an open-weight 'Wan 3.0,' a mis-dated 'Gen-4.' Treat any video-model claim without an official source or a benchmarkBenchmarkA standardized test used to compare models on a task, such as prompt adherence or image quality. entry as unconfirmed. The confirmed picture is enough.

xAI also entered in force. Grok Imagine Video 1.5, launched May 31, 2026, generates short clips with native audio and briefly topped the Artificial Analysis image-to-video arenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job., a reminder that the video race is no longer just the incumbents and the Chinese labs.