Today, we cover:
ByteDance opened public API access to Seedance 2.5 on July 16, and the number that matters is thirty. Every commercial video model before it — Google's Veo, Runway's Gen-4.5, the now-discontinued Sora — generated a clip in short native passes and stitched the rest together in post, with a shift in lighting or a drifting face at nearly every seam. Seedance 2.5 generates a continuous, unstitched 30-second clip in one inference pass, using a sparse-attention architecture that keeps the model's memory of a scene alive across the whole clip instead of a narrow sliding window. It also accepts up to fifty reference images, video clips, and audio files at once — more than triple the prior model's limit — and can edit a single region of a finished clip without regenerating the rest. None of that resolved the older problem sitting underneath it: five major studios sent ByteDance cease-and-desist letters in February over Seedance 2.0, none of the disputes have reached a US court yet, and by multiple accounts studio employees are already using the tool anyway, on what one animation producer called a "don't ask, don't tell" basis.
Three days earlier, a different bottleneck fell. Runway's Aleph 2.0, paired with OpenAI's GPT Image 2, lets an editor change a single reference frame — swap an outfit, relight a scene, replace a prop — and watch the edit propagate automatically across the rest of the clip, up to thirty seconds at 1080p, matching related shots across cuts without anyone touching them by hand. Video editing has always meant touching every frame a change affects; professional VFX studios bill by the week for exactly that labor. Reducing it to one edit and a propagation pass doesn't just speed up production — it turns a studio's back catalog of existing footage into something worth modifying rather than replacing, which is a bigger addressable job than generating new video from nothing.
Someone still has to catch what floods out of tools like these, and TikTok's answer arrived the same week. On July 10, the platform announced it had labeled more than 3 billion videos as AI-generated and taken a seat on the steering committee that writes the C2PA provenance standard, the same body that includes Adobe, Google, and OpenAI. TikTok runs three overlapping detection layers: C2PA metadata that a single screenshot strips, a proprietary watermark that survives re-encoding but only covers content made with TikTok's own tools, and automated detection that, as of late 2025, caught between 35 and 45 percent of the AI content circulating on the platform. A 2025 study from the Canadian research institute Dais tested exactly this kind of small on-screen label and found it produced no measurable change in whether people believed or shared synthetic video. The announcement lands three weeks before the EU's AI Act forces the issue anyway, on August 2.
Two of this week's stories made video easier to produce than it has ever been. The third measured, honestly, how little anyone can currently tell the difference once it's out. That gap is the actual story.
The labels arrived. The believing didn't.