Three weeks ago a client asked me to pull up Sora for a quick storyboard pass. It's gone. The web app, the ChatGPT integration, the API. All shut down over the course of this year, with the last pieces switched off on September 24. The product that started this hype cycle in early 2024 burned through roughly a million dollars a day at its peak, and still couldn't hold its user base, which dropped from about a million users to under 500K before OpenAI pulled the plug. No replacement. They're pointing people toward ChatGPT image generation instead.
Sora's death is the cleanest signal I've seen all year that the first phase of AI video is over. That phase rewarded whoever had the flashiest single clip. The phase we're in now rewards whoever can sustain a character across minutes and episodes instead of seconds. I've felt this shift on client work for months, and this week the industry data caught up to what my timelines already look like.
The moat moved from models to orchestration
Runway's pitch right now isn't "our model is better than Kling's." It's "we have Seedance 2.5, Kling 3 Pro, Gen-4.5, Nano Banana Pro and Nano Banana 2 all in one app." I stopped being loyal to any single model about a year ago. On a recent F1-adjacent brand piece I used Nano Banana Pro for a reference still, Kling for the motion pass, and Higgsfield for the fast iteration loop before final grading. No single vendor does everything well, and pretending otherwise wastes a client's budget.
If you're still picking one tool and forcing every shot through it, you're working the old way. The new workflow doesn't care which brand made the model. Learn two or three platforms deeply enough that you know which one to reach for on which shot, rather than being an expert in one and improvising badly in the rest.
Kling 4.0 wants a season, not a clip
Kling 4.0, which Kuaishou announced in late September with full rollout this month, is the clearest evidence of where the money is pointing: up to 4K output in 10-bit HDR, single generations running 30 seconds instead of 5, and up to 10 keyframes with 15 multimodal references in a single task. The reference count matters more to me than the resolution bump. Fifteen references means you can lock a character's face and outfit, plus the lighting and a prop, across multiple shots without the model drifting halfway through.
That's the problem that killed most "AI short films" over the past two years. Your lead actor subtly becomes a different person by scene four. MBC C&I and Barunson E&A, the production company behind Parasite, used Kling 4.0 at ACFM Busan for "Rendering Love," a series mixing live-action with AI virtual idols, to stress-test character consistency across episodes. A studio with Parasite's pedigree testing a model for episodic continuity, rather than a one-off ad, says something about where this is headed.
Runtime, voice, and an unlikely Oscar campaign
Higgsfield jumped from roughly one-minute video caps to generations over 10 minutes long after integrating GPT-6 Astra, according to CEO Alex Mashrabov. This is the same company that was pumping out 4 million short-form ad videos a day optimized for share velocity. Watching that engagement-optimization engine get repointed at a "Cinema Studio" workspace for trailers and short films is a strange thing to watch happen in real time. The tool that learned what makes people stop scrolling is now aimed at narrative attention spans measured in minutes. I'd treat early Cinema Studio output with some skepticism until I've seen more of it hold together past the two-minute mark, but the direction is clear.
The quieter fix is ElevenLabs' v4 model, and it's the one I'm most relieved about. I've had to rebuild entire VO passes because v3 let a single re-recorded line drift the speaker's voice just enough to be audible on a good pair of headphones. V4 holds character voices steady across a full production, supports instant clones from 10 seconds of source audio, and handles 90-plus languages with inline emotion and delivery tags. Nobody's putting this in a trailer, but it's the plumbing that keeps a 10-minute AI film from sounding like it was stitched together by three different narrators.
Then there's Fable's "Gods Don't Give Gifts," an AI-generated anthology film from YouTuber Zack London, backed by Amazon, hitting US theaters December 4 with a Best Animated Feature Oscar campaign behind it, including Academy voter screenings and an LA billboard push. Pushing the release date back from October 30 to align with awards season is a traditional-industry move for a non-traditional film. Whether or not it gets nominated, an AI film running a real campaign instead of chasing TikTok views suggests the ambitions here have changed scale.
If you're building anything longer than 30 seconds right now, here's the workflow change worth making immediately: lock your character and location references first, across every tool you'll touch, before you generate a single shot. Build a reference sheet. 3 to 5 angles, consistent lighting. And feed the same set into Kling's Omni Reference, your Higgsfield prompts, and your ElevenLabs voice clone setup. You have to architect consistency before frame one now; you can't fix it in editing.
None of this means AI film is "solved." It means the people building these tools have admitted that a good 8-second clip and a good 8-minute film are different problems, and they're finally building for the second one.