Sora is dying, not in any dramatic sense. Just a boring corporate one. OpenAI killed the web and app experiences back in April, and the API goes dark on September 24. I've had three separate client calls this month where someone asks "wait, can we still use Sora for the deliverable," and I have to explain that the tool we built half our 2025 workflow around is being quietly walked into the sea. That's a bigger deal in AI video right now than any single flashy release, because the ceiling moved, and it moved to Veo 3.1 and Kling 3.0.
The Sora exit and what actually replaces it
I've been cutting real client work on both Veo 3.1 and Kling 3.0 for the last few weeks. Veo 3.1 is the safe pick: prompt adherence is better, the native audio actually syncs, and 4K output holds up on a big screen. If a client asks what the enterprise-safe option is, I say Veo 3.1 without hesitating.
Kling 3.0 is the one I'm more excited about, though. It matches Veo on cinematic lighting and handles complex motion just as well, and the multi-shot storyboard mode with native audio sync across cuts is genuinely useful for the fast-turnaround brand and event work I do. The pricing gap isn't small either, though I'd take these numbers with a grain of salt since rates shift by tier and region: Sora 2 Pro sat around $0.75/second on the API, Veo 3.1's Fast tier runs somewhere near $0.15/second, and Kling 3.0's blended rate looks to be roughly $0.10/second. For a 30-second brand spot with several variations, that gap can eat the entire margin on the job. If a producer is watching their budget, I tell them to start with Kling.
Netflix, A24, and the money finally admitting what's happening
Netflix bought Ben Affleck's AI company InterPositive for $587 million on July 17. Affleck's framing is that it removes "logistical stuff that gets in the way" of filmmaking, which is a careful way of saying pre-vis, scheduling, and probably a chunk of post get automated. Alongside the acquisition, Netflix disclosed generative AI use across hundreds of its films and shows. Not a pilot program, not some R&D lab experiment tucked away from the business. It's already baked into the pipeline, and they're just now telling people.
A24 took a different but related path, striking a $75 million partnership with Google for AI research back in June. Their line was that they'd "rather have a seat at the table," which is the right instinct. I don't love every AI use case I've seen in film production, but I trust A24's taste here more than I'd trust a boardroom's, and having filmmakers shape the tooling beats leaving it entirely to model labs that have never sat in a color grade session.
Musk's Odyssey and the Thanksgiving Day mess
Elon Musk says Grok Imagine will produce a full-length, "historically accurate" AI adaptation of The Odyssey, while continuing to take shots at Nolan's Odyssey, a film that's doing well both critically and commercially. I'll believe the full-length claim when I see a finished cut with consistent characters across two hours, because that's still the hardest unsolved problem in AI video. Not lighting, not motion, but continuity from scene to scene. Nolan's response, that AI can't replace human creativity in filmmaking, sounds defensive until you actually sit with what "historically accurate Homer" would mean coming out of a text-to-video model trained on internet data. I suspect the vibes will be off in ways nobody quite predicts.
There's a smaller story from the same stretch that's a better warning for anyone building an AI film festival strategy. An AI short called Thanksgiving Day won the inaugural Frame Forward AI Animated Film Festival and was supposed to get a national theatrical run through Screenvision Media, which sells pre-show ads to chains like AMC. The problem: Screenvision is a pre-show ad company, not a studio or a programmer, and audiences noticed the difference immediately. Public backlash got it pulled from AMC. I don't think that means people hate AI shorts as a category. It means distribution optics matter almost as much as the work itself. If your AI film shows up in a slot people associate with concession-stand ads, you've already lost the framing battle before anyone judges the actual filmmaking.
The workflow trick actually worth stealing
Here's something I've folded into my own pipeline this month that's saved real time. Nano Banana Pro's edit-first method beats re-rolling from scratch almost every time: generate a base image, evaluate it, then give a specific note like "keep the composition but change lighting to golden hour," and refine from there instead of starting over. It supports up to 14 reference images per prompt, six high-fidelity objects, five characters. Enough that it's become the tool I reach for on multi-subject compositions where I need consistent product placement across several variations on a brand shoot.
The version I'm chaining into video work now has the model first output a full cinematic approach. Shot progression, camera movement, lens and exposure notes. Before generating a keyframe list of 9 to 12 frames, assembled into one master grid, which then feeds the video generation step. It's slower up front, but the output needs a lot less cleanup after. I'd try it before defaulting straight to Kling or Veo on the next project.
A few other things crossed my desk this week that are worth a glance, even if none of them is the headline: FLUX 3 from Black Forest Labs now does image plus 20-second audio-video from one prompt, Seedance 2.5 is bringing native 4K and 30-second outputs aimed squarely at professional advertising, and Midjourney's V8.2 preview is sharper on prompt adherence than anything they've shipped before. Any one of these could be the thing everyone's talking about in three months.
The tools keep getting better and cheaper at the same time, which almost never happens in this industry, and I don't expect it to last. At some point someone figures out how to charge accordingly.