A studio in an unnamed location just shipped a 135-minute feature film for what sounds like the budget of a nice Berlin flat. It's a full adaptation of the Odyssey, timed to land against Christopher Nolan's version, built almost entirely on Kling. I've been staring at the Fountain 0 case study for two days now, and it's changing how I think about client conversations for the rest of this year.
I run pipelines across Higgsfield, Kling, Runway, and Nano Banana Pro on paying jobs: F1 activations, Web3 summit recap films, brand spots that need to turn around in 72 hours. So when I say the tool question matters more than usual right now, it's not theoretical. Sora just proved that the ground under a production pipeline can disappear with four months' notice.
OpenAI pulled the Sora 2 consumer app on April 26, 2026. The API sticks around until September 24. After that, Sora becomes a name people cite in comparison posts instead of a tool anyone actually opens. A 30-second clip through the Sora 2 API currently runs about $22.50, roughly five times what Veo 3.1 fast mode charges for comparable output. That's a steep price for a tool on its way out the door.
Never build a pipeline where your entire post-production language depends on one vendor's roadmap. Koosha's team at Fountain 0 learned this mid-production and pivoted to Kling. The pivot worked, but plenty of teams won't be that lucky. If you've got a project mid-flight right now with Sora baked into your prompt library, migrate this week, not in September.
The independent testing community has actually converged on something useful here, which almost never happens with AI tools. Veo 3.1 wins on cinematic direction and following complicated prompts precisely. Kling wins on motion control, character consistency across shots, and cost. Kling 3.0 Standard runs about $0.84 for a 10-second clip with audio, and the output mostly holds up at that price. Veo 3.1 Lite is the cheapest option if you need enterprise infrastructure behind it, which matters more than people admit once a client's legal team starts asking about data handling.
Runway isn't playing the same game at all. It sells you an environment: keyframes, motion brush, video-to-video, camera controls, not generation by the second. For a one-off viral clip, that's overhead you don't need. For repeatable studio work where the same three motion presets have to look identical across forty deliverables for a campaign, it's the only tool built for that job. In practice: Runway for anything a client will want reshot six times with small tweaks, Kling for photorealistic movement and anything audio-led, Veo when the prompt itself is doing the heavy creative lifting and needs to be followed exactly.
ByteDance says Seedance 2.5 generates a continuous, unbroken 30-second clip in a single pass, no stitching, no visible seams. That would be a real technical jump, and nobody else has cracked it yet, but as of July 22, all that exists publicly is an announcement, a Dreamina capability page, and demo reels. Dreamina still lists it as "coming soon." No API pricing, no general access, nothing you or I can put in front of a client next week.
This keeps happening with Chinese model releases, and it's worth flagging directly: a keynote demo is not a production tool. Wait for hands-on access before you tell a client it's part of the plan. Separately, worth knowing if IP concerns ever come up in your contracts: serving legal process on a China-based company under the Hague Convention takes 18 to 24 months. That gap is a big part of why teams like Fountain 0 can move faster than Hollywood's legal department can respond, and it's not priced into most risk assessments.
Luma Ray3 generates native 16-bit HDR, exported as EXR files that drop straight into a proper color pipeline. It also handles splashing water, pouring liquids, cloth, and rigid-body collisions well enough to fool trained evaluators in blind tests. The January update, Ray3.14, added native 1080p and cut sampling cost by roughly 3x while speeding things up 40% at 720p. If you're doing anything that ends up in a DaVinci Resolve grade with real deliverable standards, and not just a social cutdown, this is the model worth testing before you default to whatever's trending.
If you're rebuilding a pipeline this month, don't pick one model and marry it. Build your prompt library and shot references so they translate across at least two generators, the way Fountain 0 ran Kling alongside Nano Banana Pro for stills and core frames. Treat the proprietary post-processing layer as the actual product. Koosha said it plainly: raw Kling output isn't what makes Odysseus watchable, the in-house tooling on top is. Whatever you build on top of the base model is the thing nobody can sunset out from under you.