What Actually Happened
The public AI video model leaderboards refreshed this month, and the top of the board is now held entirely by Chinese labs. On the text-to-video arena, Kuaishou's Kling v3 sits at number one with an arena score around 1,934, ahead of Alibaba's Happy Horse and ByteDance's Seedance, and all of its variants outrank Google's Veo 3.1 (LLM-Stats video arena; Tech Insider).
At the same time, the most famous Western product is going dark. OpenAI is discontinuing Sora, with the API endpoints for Sora 2 and Sora 2 Pro hitting a hard cutoff on September 24, 2026, and generated content deleted after that date (OpenAI). Google's Veo remains the strongest Western option, but on raw quality it now trails the Chinese frontier.
So the state of play is simple to state and awkward to act on. If you rank AI video models by output quality alone, the answer to "what is the best model" is a Chinese one, hosted by a Chinese company, running on Chinese infrastructure. For a consumer making a meme, that is a footnote. For a B2B marketing team putting client and brand assets into a pipeline, it is the entire question.
Why the Best Model Is the One You Cannot Ship
Here is the operator truth the benchmark coverage skips: in B2B, a tool is not chosen by its arena score. It is chosen by whether it survives a vendor-security review. And a model at the top of the current leaderboard is precisely the one most likely to fail that review.
Think about what a security questionnaire actually asks. Where is our data stored and processed. Who can access it. How long is it retained. Is there a SOC 2 report. Can you sign a data processing agreement under a jurisdiction our legal team recognizes. A frontier model headquartered in China, with US enterprise and government buyers already restricting Chinese-owned software as a category, struggles to answer those questions in a way that clears procurement. The precedent is not hypothetical. Enterprises spent the last few years pulling Chinese-headquartered apps out of approved-vendor lists, and that reflex now points straight at the top of the video board.
This is the part that inverts the usual advice. "Use the best model" and "pass your buyer's security review" have quietly become opposite instructions. The moment a marketing team uploads an unreleased product cut, a customer's likeness, or brand-owned footage into a top-ranked generative service, the thing that matters is not how good the render is. It is whose servers just received your client's confidential material. The leaderboard measures pixels. Your buyer measures data path. Those are different contests, and only one of them closes deals.
None of this is an accusation against any specific company. It is a description of how enterprise procurement behaves and how it is going to behave when the highest-scoring tools sit in a jurisdiction most InfoSec teams treat as a hard no. The pattern is the point, not the vendor.
The Data
Start with the third-party facts. As of this month, Chinese models (Kuaishou's Kling, Alibaba's Happy Horse, ByteDance's Seedance) hold the top of the text-to-video arena, with Kling v3 at number one and its variants outranking Google Veo 3.1 (LLM-Stats; Tech Insider). Meanwhile OpenAI's Sora API shuts off on September 24, 2026 (OpenAI). The Western flagship is leaving the field just as the ranking tilts east.
Now our first-party read. In our last twelve intake calls where a prospect asked us to build on "the best" AI video model, nine named a tool that their own company's security team would not have cleared for client-facing work, either because it is China-headquartered or because the vendor could not answer basic data-residency and retention questions. The buyers were choosing by benchmark and demo, not by what their own procurement would actually sign off on. The gap between the tool they wanted and the tool they could deploy was, in most cases, the entire top of the leaderboard.
The second first-party number is the one to sit with. Our production standard is that no generative model touches a client pipeline until it clears that client's own vendor questionnaire. In practice, applying that one rule removes every model currently ranked in the arena's top tier and leaves a short allowlist of US-hosted, enterprise-contracted tools reached through a cloud with real data terms. Even inside that cleared set, generative footage stays under roughly ten percent of finished runtime across our retainer book, and it never carries the trust moment, the founder, the customer, the exec on camera. The best-scoring model is not in our stack, and its absence has cost us nothing a client ever noticed.
The Counter-Argument, Steelmanned
The strongest case against all of this: the output is just pixels. A rendered clip does not carry a passport. If Kling produces better b-roll than anything Western, a marketer can generate it, download it, and drop it in a timeline, and no buyer inspects the render pipeline behind a finished video. On that view, worrying about the model's headquarters is theater, and the pragmatic move is to use whatever looks best and Western-clear the final file.
That case has a real edge, and anyone waving it away is not being honest. But it misses where the risk actually sits. The exposure is not the output, it is the input. To get that beautiful b-roll, you first upload a prompt, and often reference frames, brand assets, or product footage, into a foreign-hosted service that states its own retention and training terms. That upload is the event a security team blocks, and blocking it is not paranoia. It is the same logic that governs every other SaaS tool in a regulated buyer's stack. And notice the escape hatch proves the point: "just use a Western model that clears" concedes that model choice is downstream of governance, not benchmark. Add the disclosure and provenance obligations now landing under regimes like the EU AI Act, and the compliance surface only grows. The quality-first path does not remove the review. It just relocates it to the moment your client finds out where their footage went.
What To Do Monday
First, run every AI video tool through your own vendor-security questionnaire before it touches a client asset or brand-owned footage. If the vendor cannot answer data residency, retention, and processing, it is out, whatever its arena score. Treat that as a gate, not a preference.
Second, stop specifying models by leaderboard and start specifying by clearance. Keep a short written allowlist of tools that pass both your review and your buyers' reviews, and build only on those. The list will be shorter and duller than the top of the board. That is the correct outcome.
Third, keep trust assets off generative pipelines entirely. Your founder, your customers, and your executives on camera are the footage that actually moves pipeline, and they are exactly where model risk and trust risk stack on top of each other. Shoot those for real.
Fourth, if you must use generative footage, contract it through an enterprise cloud with signed data terms and cap it to non-trust b-roll. That keeps the leverage without putting confidential inputs into a service your legal team never approved.
Fifth, put the durable budget on the human-led capture and distribution system, not the model of the quarter. The system is portable across whatever tool clears review next, and it is the one asset that does not get switched off, ranked down, or blocked by a security team. Sora just proved the model is a rental. The leaderboard is proving the best rental is one you cannot sign for.