$587 million. That's what Netflix paid in cash for InterPositive, the Ben Affleck-backed AI studio, and the number only became public this week even though the deal was announced back in March. For context, that's more than Netflix spent acquiring most of its mid-tier film libraries. When a company drops that kind of money on a team whose whole pitch is generative workflows for post-production, the novelty framing stops making sense. Netflix says gen AI touched roughly 300 titles in 2026 already, mostly for crowd work and battle scenes that would've otherwise been cut for budget reasons. Nobody posts about that on X. It's still the thing actually moving money.
I spend most of my week in Kling, Higgsfield, and Veo on client work, and reading the Netflix news, what strikes me is the gap between the flashy demos everyone shares and the unglamorous technique that actually gets used on paying jobs. The demos get the attention. The technique pays the bills, and right now nowhere is that clearer than in how you talk to Kling.
Kling Now Has Dialects, and Ignoring That Wastes Credits
Kling's model family has splintered into versions that each want to be prompted differently, and treating them the same is the fastest way to burn credits on garbage generations. 2.5 Turbo Pro wants a handful of elements at most. 2.6 tolerates a longer list. 1.6 needs you to strip things down to almost nothing. Kling 3.0, the successor to 2.6, and its Omni variant (the upgrade path from O1) sit at the more forgiving end, built to handle complexity and reference-based editing without falling apart.
Here's the failure mode I see constantly, especially from people used to Sora or Veo's more forgiving text parsing: element overload. You write a gorgeous, novelistic prompt with a woman in a red coat, rain, three background pedestrians, a flickering neon sign, a dog crossing frame, and a slow dolly-in, and you feed it to 2.5 Turbo Pro or 1.6. The model doesn't gracefully simplify. It gives you warped limbs, objects merging into each other, or a shot that just ignores half your prompt.
The fix is almost embarrassingly simple, and at this point it's just a habit before I hit generate: count the nouns. Subject, action, context, style. That's roughly the skeleton, and context is where people overload things, so keep it lean if you're running 1.6 or 2.5 Turbo Pro. If you're counting more than four or five distinct things in that prompt, cut it before you submit, not after you get the broken result back.
One more distinction people get wrong constantly, myself included when I started: text-to-video needs the full scene described from scratch, but image-to-video should only describe motion. If you're animating a reference image and you redescribe the subject, the lighting, the outfit that's already sitting right there in the frame, you're adding conflicting instructions the model has to reconcile instead of clarity. Describe what moves and leave the rest alone.
Prompting Kling Like a Director Instead of a Captioner
The best change I've made with Kling this year is treating it like a camera crew instead of an image generator. Don't describe a picture. Describe a scene being filmed. "A woman in a red coat" tells the model what's in frame. "Camera pushes in slowly as a woman in a red coat crosses frame left to right, rain streaking the lens" tells it what's happening and how to shoot it, and Kling responds to that framing with noticeably more coherent motion. You've given it a job instead of a description to interpret.
It sounds like a small semantic trick, but the results don't feel small. Once you start prompting this way, going back to writing captions feels like a downgrade.
Veo 3.1's Audio Gets Good Once You Direct It
While Kling fragments into dialects, Veo 3.1 has quietly become the model to beat on sound. The technique people are sleeping on is treating audio as its own directed layer instead of a tacked-on afterthought. Don't write "with sound" at the end of your visual prompt and hope. Define the audio's purpose, source, timing, intensity, and its relationship to the camera, the same way you'd direct a sound designer on set.
Vague mood words get you generic noise. "Spooky sounds" gets you a stock horror sting that could be from any project. "Faint transformer buzz, occasional metal creak, low ventilation hum" gets you something that actually sounds like it belongs in your specific scene. And for layering, there's real vocabulary that works: foreground elements like dialogue or key sound events should be described as things that "cut through," while background elements stay audible but subordinate, described as happening "in the distance." That phrasing isn't decoration. It's how the model decides what to prioritize in the mix.
For anyone stacking models day to day, Higgsfield is worth mentioning too, not as a model itself but as the layer sitting on top of Seedance 2.0, Kling 3.0, Wan 2.6, Veo 3, Sora 2, Hailuo, Flux Kontext, and more. Its Soul ID feature, which locks a character's identity from a reference image across generations, is the closest thing I've found to solving the problem of re-prompting from scratch every single shot just to keep a face consistent.
Where the Money Actually Is
Meanwhile Fountain 0, the studio behind Tribeca's "Dreams of Violets," is back with "Odysseus: The Fall," a 135-minute fully AI-generated Odyssey adaptation with Ash Koosha. That's an ambitious swing, and I hope it gets seen. But it's not where the industry's money is actually landing. The industry's money is landing on InterPositive, on tools that make crowds bigger and battles cheaper inside otherwise conventional productions. Midjourney's V1, image-to-video only, no audio, no text-to-video, feels almost quaint next to Veo's full audio-scene approach.
What's strange isn't that AI is making feature films now. Plenty of people predicted that. It's that the $587 million bet isn't on a flashy AI-native film at all. It's on invisibility: AI work so well-integrated into a Netflix show that you'd never clock it as synthetic. That's the actual frontier, and it's a quieter, stranger place than anyone predicted two years ago.