China's AI video labs are no longer waiting for Sora and Veo to define the market. ByteDance has laid out Seedance 2.5, MiniMax is pushing Hailuo forward, and the price pressure is already visible.
If you're building video products around Western generation APIs, the uncomfortable news is not that China has one strong model. Several Chinese labs are now competing on quality, control, and cost simultaneously. That's a different kind of problem. It changes the market faster than any keynote can. ByteDance made the cleanest, best-documented move. At its Volcano Engine FORCE conference in Beijing on June 23, the company unveiled Seedance 2.5, with coverage from China AI Daily and AiRating pointing to the same core claims: a 30-second generation in one continuous segment and support for up to 50 multimodal reference inputs. The company also used the event to push native 4K support across the Seedance line. Those are not small knobs. For creators, longer clips and heavier reference control mean fewer stitched fragments and less time rebuilding a shot because the face, product, or room changed halfway through.
Access is still the caveat. Several platform write-ups in July described Seedance 2.5 as in enterprise beta or staged rollout, not a model every developer can simply open and run. You should care about that distinction. A model announced on stage is not the same thing as a model sitting inside your production workflow with stable pricing, refunds, latency, and error handling.
MiniMax is the less tidy part of the story, and frankly that matters. Hailuo H3 is showing up in late-July creator guides and MiniMax's own event materials now refer to H3 alongside M3, long-context reasoning, native multimodality, open weights, and enterprise adoption. But the hard public record does not support the cleaner claim that MiniMax shipped H3 on July 30 with a fully confirmed spec sheet and Reuters-verified pricing. That line had to go. Plausible is not good enough when you're naming an outlet.
The cheap models are already getting used #
The best evidence of the shift is not a lab claim. It is usage. Vercel's AI Gateway video leaderboard, viewed for the period ending July 28, had Seedance 2.0 Fast in first place with 25.9% of video generations, Seedance v1.5 Pro second with 21.0%, and Seedance 2.0 third with 17.8%. Veo 3.1 Fast Generate followed at 13.3%, while Kling v3.0 Image-to-Video sat at 3.8%.
That is a real signal. Developers are not only admiring Chinese video models from a distance. They are sending jobs to them.
MiniMax's public pricing page also shows why the cost discussion has teeth, even before H3 pricing is fully pinned down. A $1,000 standard MiniMax video package buys 3,760 units. Under the listed Hailuo 2.3 deductions, a 1080p six-second generation costs two units, which puts that clip a little above 50 cents before volume terms. A 768p six-second Hailuo 2.3 Fast generation costs 0.7 unit, or under 20 cents on that package math. You don't need a slogan about disruption when the invoice says enough.
The Western comparison is messier than the original draft made it sound. OpenAI's Sora is commonly tied to subscription access rather than a simple public per-second API price, while Google Veo pricing varies by tier and platform. Clean comparisons are hard. The direction is not. Chinese video models are pushing useful output toward lower unit costs, and that puts pressure on every startup trying to sell a thin wrapper around a paid generation endpoint.
The moat has moved #
This is where founders should be blunt with themselves. If your product advantage is just access to a model, you don't have much of an advantage. The model market is moving too quickly, and open-weight or low-cost Chinese releases make that weakness more obvious every month.
The stronger play is workflow. A retailer does not only need a pretty five-second clip. It needs product labels to stay readable, brand assets to remain consistent, variants to be generated in batches, and bad outputs to be rejected without burning the whole budget. A studio needs version history, review, licensing checks, and handoff into editing tools. Those are dull details. They are also where a real product lives.
ByteDance has an obvious distribution edge because it sits near CapCut, Doubao, TikTok's parent company, and Volcano Engine. The Financial Times reported this week that ByteDance has been building itself into a much broader AI platform, with Doubao reaching hundreds of millions of users in China. That is not just a model story. It is a channel story, and channels decide which tools become habits.
MiniMax does not need the same consumer footprint to be dangerous. Hailuo already has name recognition among creators, and lower generation costs can travel fast through agencies, growth teams, and small studios that test ten clips to keep one. The buyer does not care which national AI race a model belongs to. The buyer cares whether the shot works and whether the budget survives.
Independent benchmarks for Seedance 2.5 and Hailuo H3 still need to settle. Keep that in view. But the market direction is clear enough: Chinese labs are no longer arriving as cheap copies of Western video models. They are setting terms on duration, reference control, and cost. If you're building in AI video, plan around that now, not after the leaderboards make it impossible to ignore.
Also read: South Korea's Kospi posts its biggest single-day gain ever after Microsoft earnings calm AI-bubble panic; DeepSeek opens public beta API for V4-Flash as a three-way price war hits AI developers; Kioxia's 31-fold profit surge still wasn't enough to satisfy Wall Street