AI Video Generation Pivots to Native Audio After Sora Shutdown

A market that changed shape in six months

By mid-October 2026, generating a video from text works differently than it did at the start of the year. The best-known name in the category is gone. OpenAI closed the Sora app and website on April 26 and switched off the developer API on September 24, ending a product it first demoed in February 202421. Over the same stretch, Kuaishou, ByteDance, Alibaba, MiniMax, Google and Runway shipped models that produce sound with the picture, accept more reference material, and increasingly edit footage rather than just generate it23.

The main story is not which model leads a leaderboard this week. Taken together, the coverage shows AI video becoming one piece of a production pipeline. It now sits alongside image generation, voice synthesis and general multimodal models, and creative teams choose tools by how well they fit that pipeline, not by how impressive a demo looks.

Sora's exit and the lesson about availability

The reports agree on the shutdown dates. OpenAI announced the decision on March 24, closed consumer access in April, and ended the API in September2428. OpenAI's deprecation page names no replacement, and the company did not move video generation into ChatGPT2821. The explanations differ by outlet. One account points to the compute the video product used, capacity other teams inside OpenAI needed30. Another cites media reports of high computing costs, fading interest and rights disputes21. A third notes that a reported Disney licensing deal ended as well30.

The takeaway for working creators is that a strong model does not guarantee a stable product. One comparison makes the point directly: a filmmaker planning a six-month production needs the platform to still exist when the project ends, however good the underlying model is26. Buyer's guides now describe Sora as something to migrate away from, not a tool to build on27. Migration guides also warn that apps still selling "Sora 2" after April are either using another model under the old name or running a scam21.

Native audio is now standard

The clearest technical trend of 2026 is video and sound generated in a single pass. Kling 3.0 arrived in February with clips up to 15 seconds, native audio and speech in five languages, all produced together so dialogue and effects stay in sync1. ByteDance's Seedance 2.0 followed days later, accepting mixed text, image, audio and video inputs2. Its July successor, Seedance 2.5, doubled clip length to 30 seconds and added multi-round extension24.

One comparison argues that native audio now matters more than raw visual quality when choosing a tool6. The open-weight community is moving the same way. Kandinsky Lab released Kandinsky 6.0 Video under an MIT license on October 6. It generates five-second clips with synchronized 44 kHz audio, including lip-sync136. Lightricks' LTX-2.5, released in August, offers text-to-video, image-to-video and audio-to-video endpoints at up to 4K2.

Runway is the notable exception. One detailed comparison finds that Runway's official documentation for Gen-4.5 does not list native audio. Instead, Runway sends users to separate audio and post-production tools26. A pricing table backs this up, showing no audio in Gen-4.5 through the API28.

Editing tools are where creators gain the most

Runway's response is to lean into editing. Aleph 2, released in May, restyles and edits existing footage from text prompts and runs in a panel inside Adobe Premiere Pro and After Effects1. Ruby, launched August 20, converts standard-dynamic-range video to HDR and exports in professional formats such as ProRes 422 HQ and 16-bit EXR5. Luma pursued professional color work from a different angle. Its Ray3.2 model, released in June, accepts up to 16 keyframes, generates 20-second 1080p clips and keeps native 16-bit HDR EXR output1.

This is the strongest argument that AI video is becoming useful to professionals. Keyframes, reference images, HDR export and integration with editing software let a model slot into an existing pipeline instead of replacing it. One analysis concludes that the best video model and the best video product are no longer the same thing26. The evidence supports that. Google's Gemini Omni Flash, its recommended default for video generation, is built around conversational editing: users generate a clip and then refine it with follow-up prompts1.

The drawback is clip length. Most single generations still run about 5 to 15 seconds, so longer pieces are assembled shot by shot24. Veo 3.1 generates eight seconds per pass but can be extended up to 20 times, reaching 148 seconds at 720p29.

The leaderboards disagree

Anyone hoping for one clear ranking will be frustrated. One October roundup puts Alibaba's Wan 3.0 first on the Artificial Analysis text-to-video board, both with and without audio1. Another, published two days later, says Gemini Omni Flash leads that same arena7. A third describes the top of the audio-enabled table as shared among Wan 3.0, Gemini Omni Flash, MiniMax H3 and Seedance 2.0, with Veo 3.1 down in the teens30.

Smaller details conflict too. Coverage of Kling 4.0 Flash describes early-access clips of 3 to 30 seconds at up to 4K in one account1, and up to 20 seconds at 720p for Ultra yearly subscribers in another3. Reports of Seedance 2.5's reference limit range from 50 combined inputs3 to 30 images, 10 videos and 10 audio clips8. These gaps likely come from rapid version churn and vendor-supplied specs, and several outlets label performance figures as self-reported4. The sensible conclusion is that the top tier is tightly bunched, so price, control and availability should drive decisions more than small Elo differences.

Price is where the field differs most. MiniMax H3 is described as scoring within a few points of the leader while costing less than a third of typical per-second rates at 2K7. Its weights are open, and community tools now run a trimmed version on GPUs with as little as 8 GB of memory236.

Voice synthesis grows into a performance tool

Voice is moving in parallel. ElevenLabs released Eleven v4 and v4 Turbo on September 28. The models support more than 90 languages, up from about 70, and can clone a voice from 10 seconds of audio1519. One analysis presents the release as a shift from text-to-speech that reads words aloud to models that work out how a line should be performed from character and scene context. It also notes that v4 ranked first on a third-party voice arena with an Elo of 131912. The same report cautions that ElevenLabs' claim of roughly 75% listener preference comes from an internal test with no published methodology12.

For video work, the most relevant features are stackable inline direction tags, steadier voice identity across long-form and multi-speaker content, and v4 Turbo's reported 150-millisecond time to first speech151912. The business numbers suggest strong demand. A Reuters report says a $300 million employee tender valued ElevenLabs at $22 billion, double its February valuation16. The legal risks are also real. A class action filed in Illinois in May alleges voice data was collected without required consent, and Eleven v4 does not resolve it12.

Image and multimodal models feed the video pipeline

Image generation increasingly serves as the planning stage for video. Google made Nano Banana 2.1 generally available on October 6. It brings better character consistency and text rendering, and adds panoramic aspect ratios up to 8:1 at resolutions up to 4K. Black Forest Labs released FLUX 3 Image on October 1, with bounding-box layout control, up to ten reference images and native 4K output3135. That model belongs to a wider FLUX 3 family that also generates video with synchronized audio131. Google's API can even work in reverse, taking a video as input to produce thumbnails or posters.

At the research frontier, Reka's Rho-1 preview aims to understand and generate text, images and video in one model. Reka itself calls it a proof of concept, not a finished product35.

The verdict

Read together, these reports point to a clear conclusion. AI video in late 2026 is not a contest to replace production crews with a text prompt. It is a modular toolkit: image models for consistent frames, video models with built-in sound or strong editing, and voice models for performance and localization. The winners are the vendors that fit into existing pipelines and stay available. Sora's exit is the clearest warning about the second requirement. For creative teams, the practical approach is to plan for switching between providers, prefer tools with export formats and integrations they can rely on, and treat any single leaderboard score as a snapshot rather than a final ranking.