On July 31, something interesting happened in the AI video world: ByteDance's Seedance 2.5 and MiniMax's Hailuo 3.0 (H3) launched on the exact same day. It's no coincidence — it's a signal of how white-hot the race has become. Neither side wanted to let the other claim the "next-gen video model" spotlight alone.

The most eye-catching upgrade in Seedance 2.5 isn't just that duration has stretched to 30 seconds — it's native audio. Video and sound are generated in one pass, no post-production dubbing required. Add to that 50 multimodal reference inputs, precise local editing, and stronger temporal consistency, and the competitive baseline for Chinese AI video models has been raised an entire notch.

From Second-Long Demos to 30s with Native Audio: The Bar Jumps

AI video competition back in 2024 was still about "who can generate a few more realistic seconds." By 2025, it was about duration and consistency. In 2026, the dimensions of competition have completely changed.

Seedance 2.5's key upgrades all point toward one core question — can it be used in actual production workflows:

30-second single-clause native video — going from 5s to 10s to 30s isn't just a duration increase; it's a qualitative leap in scenario usability. Short ads, social media content, product demo videos — 30 seconds is the baseline threshold for something you can actually use. Before, AI-generated video could only make the opening few seconds of a clip. Now you can make a complete short-form piece from start to finish.

Native audio generation — this is what many people overlook, but it's extremely important. Video isn't just visuals; there's voiceover, sound effects, background music. Before, after generating video with AI you still had to find TTS for dubbing, find music assets, do sound design — go through the whole workflow and you don't actually save that much time. Native audio makes the whole pipeline end-to-end, meaning one person can truly produce a complete, usable short video from just a text description.

50 multimodal reference inputs — you can simultaneously reference images, videos, characters, styles and up to 50 sources to precisely control the output. This is huge for professional creators — it's not about letting AI make things up randomly; it's you providing references and constraints, and AI works within your framework. Creative control comes back to the human.

Precise local editing — edit a region of the video, a certain shot, a segment of time, without regenerating the whole thing. This is the key step from "one-shot generation" to "iterative creation" — same logic as layers in Photoshop and timelines in editing software.

While Sora's Been Missing, Chinese Players Are Defining the Baseline

It's a fascinating phenomenon: Sora, once considered the "ceiling" of AI video, now feels like it's just… gone.

More than two years after its announcement, regular users still can't use it. No pricing, no public API, no creator ecosystem. Just occasional demos and rumors reminding everyone Sora still exists. But during Sora's absence, Chinese AI video companies haven't been idle — Seedance, Kling, MiniMax's Hailuo, plus a whole ecosystem of aggregation platforms and vertical apps, have already built out the AI video industry.

More importantly, Chinese players are now defining the "competitive baseline" of AI video. Before, everyone benchmarked against Sora's demo quality. Now, people compare: who has longer duration, who supports native audio, whose editing features are stronger, whose API is cheaper and more stable, who has more ecosystem partners. These aren't flashy lab demo metrics — these are the hard metrics of real product competition.

Of course, this doesn't mean Chinese companies have surpassed Sora technically. In ultimate image quality, complex physics simulation, and logical consistency over long videos, Sora probably still has an edge. But technical advantages that don't translate into products and ecosystems shrink over time — just like in the early search engine industry, the best technology didn't always win in the end.

After Seedance 2.5, the AI video war enters a new phase: from "who has the prettiest demo" to "who has the most usable product." Thirty seconds, native audio, editability, multimodal references — put these capabilities together, and it means AI video is finally moving from "toy" to "tool," from "creator showboating" to "content production infrastructure."

When more and more short videos, ads, and e-commerce content start being generated with AI, when creators' workflows truly get restructured by AI — looking back, maybe we'll realize that the real inflection point for AI video wasn't one amazing demo. It was a bunch of "good enough but imperfect" features piling up, until suddenly the threshold of usability was crossed.

See you tomorrow.