OpenAI is done with Sora, and the three companies that filled the gap it left behind now split AI video into three separate jobs: ByteDance handles the physics, Runway holds the story together, and Luma hands the file to your colorist.
Sora's death was slow, then fast. OpenAI pulled the plug on the Sora web and mobile apps on April 26, 2026, and the API version follows on September 24, just days from now. The math behind the call was brutal. Sora reportedly cost OpenAI close to $1 million a day to run while pulling in only about $2.1 million in total revenue over its life, and its monthly active users had slid from roughly a million at launch to under 500,000 by the time the shutdown was announced. A collapsed licensing deal with Disney and a mounting pile of deepfake complaints didn't help. OpenAI wanted a cleaner balance sheet ahead of its IPO, and Sora cost more than it made.
That left a real gap. It didn't stay empty long.
Three companies, three jobs #
ByteDance moved first with Seedance 2.5, which its Seed research team launched on July 31, 2026, rolling it out through Jimeng AI and Doubao Pro before opening API access via BytePlus ModelArk. The pitch is control over physics, not just pixels. Seedance 2.5 generates native 1080p video up to 30 seconds long, with audio built in, and it accepts as many as 50 separate reference files, a mix of images, video, or audio, that a director can feed in to lock down lighting, facial expression, and motion paths before a single frame renders. It also edits in place. If a character's hair is the wrong color three seconds into a clip, you fix that patch of the frame instead of regenerating the whole 30 seconds and hoping the rest still matches.
Seedance wins the single-shot argument. Runway wins the one that spans several shots at once. Gen-4.5, the company's refinement of the Gen-4 architecture that introduced reference-based consistency, targets the problem that has quietly sunk more AI film projects than bad rendering ever did: a character who looks like a slightly different person in shot four than in shot one. Gen-4.5 extracts visual traits from an uploaded reference and holds them across a multi-shot sequence, now supporting character-consistent generation up to a full minute, with native audio and multi-shot sequencing layered on top. For anyone building an actual narrative rather than one flashy clip, that consistency is the whole product.
TikTok Founder Zhang Yiming Is Now Asia's Richest Person, Beating Adani ByteDance founder Zhang Yiming has overtaken Gautam Adani to become Asia's richest person, with a net worth above $105 billion, according to the Bloomberg Billionaires Index. The surge, an eightfold jump from $13 billion in 2019, is being driven by ByteDance's Doubao AI chatbot and a $550 billion valuation, not TikTok ad revenue. - Zhang Yiming net worth surpasses Gautam Adani Asia - ByteDance founder becomes Asia's richest person AI wealth
Luma took the third lane, and it's the least glamorous of the three. That's the point. Ray 3.2, released June 9, 2026, doesn't try to out-render Seedance or out-act Runway. It exports native 16-bit HDR footage as EXR files in the ACES2065-1 color space, the same format professional colorists already expect from a RED or Alexa camera. That's the format a colorist can actually use. An AI-generated shot can drop straight into DaVinci Resolve or a traditional VFX pipeline without the dynamic range getting crushed on the way in. That used to keep AI video out of real productions.
One tool per job #
Here's the thing that actually matters if you're running a production pipeline this month: you no longer pick one AI video tool and live with its compromises. You pick a tool per job. A single ad might get its hero shot from Seedance for the reference-driven realism, get intercut through Gen-4.5 to keep a spokesperson's face consistent across five setups, then get graded and finished through Ray 3.2's HDR export so the color team never has to touch it twice.
That's a genuine shift from where this category sat a year ago, when Sora, Gen-3, and early Kling releases were still mostly competing over who could fake the most convincing six-second clip. Nobody's chasing novelty anymore. Seedance, Runway, and Luma are each betting that the real money in AI video sits in the parts of a production pipeline that used to require a full crew: the reference wrangling, the continuity supervisor, the colorist. OpenAI built the demo. It never built the pipeline.
Also read: Gavin Newsom Orders California To Build The AI Kill Switch He Once Vetoed • Huawei Speeds Up Its Ascend 960 AI Chip Launch to Take On Nvidia in China • Yoshua Bengio Warns AI Needs Nuclear-Style Guardrails Before It's Too Late
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.
ByteDance Signed a $29.6 Billion Loan to Fund Its AI Buildout Abroad ByteDance closed a $29.6 billion syndicated loan with 28 banks, according to Bloomberg, making it Asia's second-largest dollar loan this year behind SoftBank's $40 billion OpenAI-linked facility. Chinese lenders covered 64% of the total, led by ICBC, Bank of China, HSBC and China Construction Bank, as ByteDance funds AI infrastructure and data... - bytedance 29.6 billion dollar loan for ai - largest syndicated loans in asia this year