AI Video Generation Shifts East as Kling 4.0 and Wan 3.0 Lead
Kling 4.0 has launched with 30-second clips and Wan 3.0 leads the benchmarks. With Sora shut down, Google shifts video to Gemini Omni as AI audio becomes
Multimodal AI models are systems built to understand and generate across more than one type of data at once—text, images, video, and audio—rather than being confined to a single format. Instead of a chatbot that only reads and writes words, a multimodal model can watch a video and describe it, turn a written prompt into a moving scene, or summarize notes into spoken audio. This flexibility is what's driving the technology's rapid expansion into consumer products, from voice assistants on phones to note-taking apps that now generate video recaps automatically.
The pace of development has accelerated sharply as major labs and platforms race to outdo one another on realism, speed, and reasoning ability. Video generation in particular has become a proving ground, with new models pushing toward longer, more coherent, and more controllable outputs. At the same time, these same generative capabilities have made it trivially easy to produce convincing fabricated footage, raising fresh concerns about misinformation, political manipulation, and the erosion of trust in visual media—concerns that have already surfaced in viral, deliberately misleading AI-generated clips.
This hub tracks the full arc of that story: new model releases and benchmark claims from major AI companies, the integration of multimodal capabilities into everyday consumer tools like phones and productivity apps, the competitive dynamics shaping who leads in video and voice generation, and the growing debate over misuse, authenticity, and regulation. Readers will also find coverage of the infrastructure and economic pressures—compute costs, energy demands, and investment trends—that underpin this fast-moving corner of the AI industry, along with analysis of how multimodal systems are reshaping everyday interactions with technology.
Kling 4.0 has launched with 30-second clips and Wan 3.0 leads the benchmarks. With Sora shut down, Google shifts video to Gemini Omni as AI audio becomes
Kling 4.0 launched with 30-second clips and 15 references as FLUX 3 Klein, MOSS-Image, Voxtral 3 and Muse Spark Open pushed multimodal AI toward structured
VideoGen launched Model 1, its first in-house AI video model, which makes multi-shot videos up to 10 minutes long, plus a $1M credit fund for creators.
Higgsfield says OpenAI's GPT-6 Astra lets its agents produce AI videos over 10 minutes, up from about one minute, with voice-driven filmmaking planned next.
One2X's Medeo expanded into a Creative OS that brings AI video trend discovery, creation, social publishing and a creator marketplace into one workflow.
VSDC 11.3 adds in-editor AI video and image generation via fal.ai, plus Garmin heart rate, cadence and compass data, DJI telemetry overlays and new templates.
A report says Seedance 3.0 launched with 30-second clips and 50 references, but those specs match ByteDance's Seedance 2.5, and no official 3.0 is confirmed.
ElevenLabs reached a $22B valuation via a $300M tender offer and launched Eleven v4 as voice AI moves into enterprise agents, acquisitions and video models.
AI video tools like Kling 3.0, Runway and Vidu Q4 now use reference images, voice binding and multi-shot storyboards to keep characters consistent across
Eloelo Group unveiled Dolphin AI at FICCI FRAMES 2026, an agentic, model-agnostic video studio promising shot-to-shot continuity, now in early access.
Anthropic launched Claude Dashboards and Motion in beta on Oct. 8. Motion builds editable explainer videos from code rather than using a text-to-video model.
HiDream.ai released vivago R1 globally, an agent-based tool that scripts, storyboards, generates and edits videos up to five minutes long, with a CLI for
Shengshu released Vidu Q4 Preview on Oct 7, 2026, adding 15 image references, 3 voice references, native audio and 4K output, debuting third on an I2V
Google, ByteDance and Kling launched cheaper, longer AI video models, OpenAI and Canva pushed image editing, and ElevenLabs and AssemblyAI released voice
OpenAI shut down Sora as Kling 4.0 launched, Canva released its Craft image model, and ElevenLabs, Mistral and AssemblyAI shipped new voice models.
OpenAI's ChatGPT Images 2.0 (GPT Image 2) can search the web, plan images before drawing them, and make up to eight consistent images per prompt.
A Discover Informatics review of 120 studies covers 13 multimodal AI models from CLIP to Llama 4, finding strong benchmark results but persistent hallucination.
OpenAI's December 2025 ChatGPT Images launch started a run of releases up to Images 2.5, while Sora video shut down and the Realtime voice models grew.
Kuaishou released Kling 4.0 globally with 30-second clips, 15-reference control and keyframes as AI video and voice model rivals multiply.
Alibaba's Wan 3.0 and Google's Gemini Omni Flash now lead text-to-video rankings as OpenAI's Sora shuts down and native audio becomes standard.
OpenAI shut down Sora as Google, ByteDance and Alibaba race ahead with native-audio, long-clip AI video models that are rewriting how creators produce video.
AI video, image and voice tools are merging into unified multimodal platforms, reshaping the competition among OpenAI, Google, Adobe and Runway.
After OpenAI shut Sora down, AI video rivals moved to built-in sound and editing tools, while ElevenLabs v4 and new image models filled out creative pipelines.
OpenAI's ChatGPT Images 2.5 adds a Sketch tool, faster generation, and precision editing; here's what's confirmed versus claimed across the coverage.
Maggie Gyllenhaal scrapped AI face-replacement tech in her Marilyn Monroe short after it undercut Dakota Johnson's performance, she said in Venice.
A developer built a real diffusion transformer that generates AI face images on a $1 RP2350 microcontroller in 5 to 20 seconds per image.
Open text-to-video models on Hugging Face now generate audio: MiniMax H3 and LTX-2.5 add synced sound, while Wan 2.2 stays the top Apache 2.0 pick.
Adobe adds a Markup tool to Photoshop for drawing prompts, part of a wider AI image-generation surge across ChatGPT, Google, and social media.
LM Studio added Z.ai's GLM-5.3-Flash to its Bionic agent app on Aug 26, with image input, 1M-token context and cloud pricing about 10x cheaper than GLM-5.2.
Alibaba's Qwen released Qwen3.8-Flash, a 125B multimodal MoE model trained on about one-ninth the resources of Qwen3.7-Plus, previewing Qwen4.