Last September, every creative director I know was obsessed with getting Sora access. This week, Sora is gone. Not "deprioritized," not "sunset gracefully." The app shut down April 26, the API died September 24, and I just watched a tool directory quietly replace every Sora mention with "try Gemini Omni Flash, MiniMax H3, or Kling 4.0 instead." Twelve months. That's the entire lifespan of what people were calling the GPT-3.5 moment for video. I don't think there's a faster flagship-to-obsolete cycle in the history of generative media, and I say that as someone who's lived through three different "best model" pivots in my own project files this year alone.
I'm writing this from Berlin, mid-project on two client deliverables that both route through Kling at some point in the pipeline. So forgive me if this reads less like analysis and more like a field report, because that's what it is.
The Power Shift Nobody in the US Wanted to Admit Was Happening
Here's the uncomfortable part for anyone who built their workflow around American tools over the past two years. The leaderboards right now are a Chinese sweep. Kling v3 leads text-to-video arena scores with 1934, trailed by Happy Horse at 1816 and Seedance 2.0 Fast at 1747. Scroll down the "best models this month" lists and you'll find Wan 3.0, Seedance 2.5, HappyHorse, and Kling occupying most of the top tier, with Veo 3.1 and Runway holding ground but no longer dominating.
And then Kling 4.0 dropped on September 28, and it's not a small update. 4K output in 10-bit HDR, single generations up to 30 seconds, up to 10 keyframes, 15 multimodal references in one task. I ran the Flash preview tier on a test clip this week (it's rolling out to limited creators right now), and the temporal consistency on a 20-second shot with a moving subject was better than anything I got out of Sora 2 in its final months. Kling is sitting on more than 60 million users and 600 million generated videos. That's not a lab experiment anymore. That's infrastructure.
If you're still building your toolchain assuming American labs stay ahead by default, you're planning for a market that stopped existing sometime around April.
Studios Are Starting to Act Like AI Studios, Not AI Departments
The more interesting shift, honestly, sits upstream of any single model. Utopai Studios in Mountain View just launched its own proprietary model, Utopai X, which landed at No. 2 on Artificial Analysis' Text-to-Video Leaderboard with Audio, the highest-ranked US model on that board. They've got 25 original projects moving through development and production, a billion-dollar valuation, and they grew out of a company called Cybever that nobody outside this space had heard of two years ago. This isn't a plugin. This is a studio that happens to run on a model it built itself.
Lionsgate took a stake in Runway back in June. Google invested in A24 days later. Netflix signed a reported $600 million AI-augmented filmmaking partnership, and Warner Bros. Discovery says roughly 40% of its unscripted content now touches AI somewhere in the pipeline. Analysts are throwing around numbers like 30% drops in production budgets and 30% shorter post timelines, mostly from automating color correction, stabilization, and rough cuts.
I believe the budget numbers. I'm less sold on the framing that this is primarily about replacement. An Utopai executive put it well at Karlovy Vary this week: something along the lines of AI being less about replacing artists and more about restoring a scale of filmmaking that mostly vanished from Hollywood over the last two decades. That tracks with what I see on client work. Nobody's asking me to replace a director. They're asking me to get them three more ambitious concepts into the same three-week window, because the budget for extra shoot days simply doesn't exist anymore.
Small studios are reading the room the same way. AiCandy in Sydney, co-founded by Kent Boswell and Marcus Tesoriero, is explicitly selling "hybrid human plus AI" rather than full automation, and their pitch is basically: we came from story and craft first, tech second, and that ordering matters. I'd bet on that positioning outlasting the pure-automation pitch decks.
The Ad World Is Where You Actually See the Receipts
Forget the think pieces. Watch what's shipping. Popeyes put out "Wrap Battle," a fully AI-generated diss track aimed at McDonald's, produced by PJ Accetturo entirely in Google Veo 3 with Suno handling the music, turned around in three days. It went properly viral across TikTok, Instagram and X. Three days. I've had single VFX shots take longer than that.
Under Armour's "Forever Is Made Now" campaign for Anthony Joshua is the more instructive case for anyone doing brand work. No direct access to the athlete. They built it from existing assets and a 3D model, stacking AI video, AI photo, 3D CGI, traditional VFX, 35mm film, and AI voiceover into one coherent three-week deliverable. That's the actual skill right now: not prompting a single tool well, but knowing which five tools to stitch together and in what order. The ethics-and-crediting debate around that piece is real and unresolved, and I'd rather see it argued out in public than quietly ignored.
Then there's the cautionary tale. A Seedance 2.0 deepfake clip of "Tom Cruise" fighting "Brad Pitt" went viral earlier this year and got the MPA to publicly denounce it as massive infringement. That incident is exactly why Kling and Veo's more cautious, licensed approach to likeness and character cameos exists. If you're doing brand or talent work right now, assume every platform is tightening likeness controls, not loosening them, and build your releases and model consent paperwork accordingly before a client asks.
Actionable note if you're producing on a deadline this quarter: stop defaulting to whatever model you used last project. Run your opening shot through Kling 4.0 Flash, Seedance 2.5, and Veo 3.1 in parallel before committing to a pipeline. The gap between them on motion consistency and prompt adherence is wide enough right now that picking wrong costs you real revision time. Thirty minutes of comparison testing up front saves a day of regeneration later.
Hollywood Still Doesn't Trust Any of This, and That's Fine
Luca Guadagnino's "Artificial," starring Andrew Garfield, premiered at NYFF on October 5. It's a satire about the origins of the AI boom, and early reactions are praising its biting tone. It's landing in the same season where Variety is running headlines like "there is nothing like human," pointing at audience backlash against AI-made films even as studios quietly pour money into exactly that. I find that tension healthy rather than contradictory. The industry can build the infrastructure and still be nervous about what it's building. That's usually how it goes with tools that actually change the economics of a craft.
What I keep coming back to is this: the models will keep churning. Kling 5 or whatever Chinese lab leapfrogs next will show up in six months and make this whole post read quaint. The thing that won't change is that the people getting real work made well are the ones treating these tools as part of a production pipeline with judgment attached, not as a replacement for having a point of view in the first place.