Sora burned $15 million a day and made $2.1 million total. Disney had a billion dollars on the table for that deal and walked away. OpenAI is killing the Sora API on September 24; the web and app versions already went dark back in April, and barely anyone seemed to notice.
The interesting part isn't that Sora died. It's what was happening while it burned money on dancing avocados for Twitter engagement. Three or four other labs quietly built tools people use on paying jobs, and most of that activity isn't showing up in Western trade press.
The leaderboard tells you something different
Kling v3 is sitting at the top of the current blind-vote arena with a score of 1991, ahead of Happy Horse 1.0 and Seedance 2.0 Fast. The ranking comes from 1,328 blind comparisons across text-to-video, image-to-video, and editing, where the voter has no idea which model made which clip. The gap between Kling and what Western labs shipped over the last year shows up clearly across those comparisons.
ByteDance also dropped Seedance 2.5 into late beta. For anyone doing previz work, it's worth a look: native 4K, 30-second outputs, and a 3D pre-visualization layer that lets you block camera moves and lighting on a rough virtual set before committing to final render. Liang Rubo said it was built with film directors' workflows in mind, and the feature set backs that up. It's the first AI video tool that seems to have copied something a working DP or director actually does on set, rather than what a prompt-engineering demo assumes filmmaking looks like.
Workflow note: if you're testing Seedance 2.5 when it opens up, block your scene in the 3D previz mode first, lock camera moves and rough lighting there, and only then push to final texture render. Treat it like a virtual production stage, not a text box. You'll waste far fewer generations chasing a shot that was never going to work at the blocking stage anyway.
Tilly Norwood is getting a feature film, and the union is not having it
Particle6 announced "Misaligned," a comedy-drama starring Tilly Norwood, the AI actor that's been generating outrage since last year. The production model is worth noting: real directors, writers, and editors working alongside AI specialists, with mentorship built into the process rather than AI replacing the crew wholesale. That's a more specific structure than most AI film announcements bother laying out.
SAG-AFTRA's statement was blunt: Tilly Norwood "has no life experience to draw from, no emotion," and audiences aren't interested in content "untethered from the human experience." That's a reasonable instinct, though animated and CG performances have had zero life experience behind them for decades and audiences loved them anyway. Whether Tilly is "alive" probably matters less than whether "Misaligned" is a movie anyone wants to watch twice. A harder problem for Particle6 than any union statement.
Guerrilla filmmakers are outpacing the studios
The most culturally loud AI video of the week wasn't from a studio at all. It was Charlie Curran's Batman-inspired attack ad for LA mayoral candidate Spencer Pratt, depicting the city as run by cake-eating, vodka-chugging aristocrats in Joker makeup. Tasteless, funny, and it went viral off the back of tens of millions of views he's already built since January. He made it with Seedance 2.0, a Chinese tool most American political consultants haven't touched.
This part of the AI video story keeps getting missed by people who only track OpenAI and Google announcements. The most-watched AI video content of 2026 isn't coming from studio pipelines or ad agencies. It's coming from one guy with a laptop and a Chinese model nobody in his industry has vetted for bias, provenance, or watermarking. Political operatives should probably be paying more attention to that than they currently are.
Money and infrastructure catching up to the hype
PixVerse just closed a Series C extension at $439 million raised total, crossing a $2 billion valuation, with a new V-Series model and world model on the way this year. That's real money chasing a company that, unlike Sora, has retention numbers behind it.
Meanwhile Google quietly shipped something that may matter more long-term than any single model release: Google Vids now lets you build a custom avatar of yourself from a selfie and a voice clip, tied to your account, invisibly watermarked with SynthID. That's a different instinct from the Sora era, where anyone could generate anyone doing anything and sort out provenance later. Here, provenance is built in from the start rather than bolted on after a controversy. A less flashy decision, but probably the right one.
If you're building an avatar pipeline right now, HeyGen just added ElevenLabs V3 support to its Create Avatar Video API, alongside the existing multilingual and turbo models. It's a small changelog line, but a real workflow upgrade if you're stitching voice and avatar tools together for client deliverables.
Sora's death isn't the end of AI video. It's the end of pretending burn rate was a strategy. The tools that survive from here will likely be the ones that make money, solve a real production problem like previz, or build in enough of a paper trail to survive regulation. Everything else was a demo reel.