The first time you got a usable clip out of Sora, it probably felt like cheating. A camera move that would have cost you a dolly track, a rain-soaked street that didn’t exist, a character who stayed consistent from shot to shot. Then the plug got pulled, and a folder full of half-finished storyboards suddenly had nowhere to go.
If you’re mid-project right now, you’re not starting from scratch. You still have the thing that took the longest to build: a shot list and a set of prompts that have already been debugged. Here’s how to move a real project off Sora without losing a fortnight.
Step 1: Rescue your archive in the next 48 hours
Download first, decide later. Old dashboards have a way of quietly disappearing, and the version of a clip you shared to Slack is not the version you want to grade and deliver.
- Grab the highest-bitrate file available for every clip you might use, not just the winners. That “failed” take is often one trim away from being the establishing shot you need.
- Copy your prompt text into your own notes. A prompt history inside a web player is a convenience feature, not a backup. Plain text files in a folder called video-prompts will outlive every model you’ll ever pay for.
- Note which clips carry a visible watermark or a licence you can’t use commercially. Flag those now, while you remember which client they were for.
- Mirror everything to two locations. An external drive plus cloud storage is enough. Editing off a single drive is how people lose a week.
While you’re at it, a quick reality check on why the tool vanished helps you plan around the next one: the write-up on what actually killed OpenAI’s viral AI video model lays out the commercial pressure that ended it, and the thinking behind OpenAI’s decision is worth ten minutes if you’re about to bet another project on one vendor.
Step 2: Audit your shots by what they actually need
Most people jump straight to shopping for a new tool. Don’t. Spend an hour listing every shot and the one thing that makes it work. A talking head needs lip sync. A product turntable needs consistent geometry. A drone-style reveal needs a smooth, believable camera path.
Sora was strong on extended takes with coherent motion and a certain filmic sheen. It struggled with hands, small on-screen text and anything involving a ball bouncing off a wall. Score your list honestly against those strengths and you’ll find that maybe a third of your shots don’t need a generative model at all. A stock clip, a £40 prop and a phone on a gimbal will beat a week of prompt iteration on a five-second insert.
Step 3: Match each shot type to a replacement model
There’s no single heir to Sora, and that’s fine. Route by job:
- Dialogue and anything that needs sound: Veo 3 generates audio alongside the picture, which removes an entire foley pass from your schedule.
- Character continuity across multiple shots: Runway’s reference-image workflow and Kling both let you carry a face or an outfit from shot to shot with reasonable fidelity.
- Fast social iterations: Pika and Luma Dream Machine are cheap to hammer through twenty variations in an afternoon.
- Animating a still you already like: Kling and MiniMax’s Hailuo are the usual picks for image-to-video.
- Self-hosted or privacy-sensitive work: Wan and HunyuanVideo under open licences, if you have the GPU budget.
Your character sheet matters more than your model choice. Lock a reference with two or three angles, consistent lighting and neutral expression. If you need help generating that sheet, the approach behind OpenAI’s newer image model is a good template for producing reference frames that video tools will actually accept.
Step 4: Rewrite your prompts as shot specifications
Here’s the rewrite that trips most people up. A Sora-era prompt looked like this:
“A woman in a red coat walks through a rainy Tokyo alley at night, cinematic, 4k, masterpiece, moody.”
Fourteen words of vibe. On newer models you’ll get better results from something closer to a camera sheet:
“Medium tracking shot moving right to left with the subject. Woman in her thirties, dark shoulder-length hair, wet red wool coat, walking away from camera at a steady pace. Narrow alley, wet asphalt reflecting neon signage, heavy rain, shallow depth of field, magenta and warm amber light sources from shopfronts. Eight seconds, 24fps.”
The differences that matter: name the camera move and its direction, give your subject three fixed attributes, state where the light is coming from, and specify duration and frame rate. Drop the quality incantations. “4K, masterpiece, best quality” does almost nothing on modern models and often adds haze or over-smoothing.
Keep a prompt template
Subject and wardrobe. Action and pace. Camera move and direction. Location and weather. Light sources and colour. Duration and frame rate. Fill all six lines every time and your hit rate roughly doubles, regardless of which model you’re pointing at.
Step 5: Rebuild continuity, then expect carnage
Continuity used to be Sora’s party trick. Now you engineer it. Use first-frame and last-frame conditioning so the end of shot one matches the start of shot two. Keep a seed log, because the seed is the only cheap variable you control. Lock wardrobe, hair and lighting in writing before you generate anything, then stop changing them halfway through because a shot looked slightly dull.
Budget for waste. On a well-specified shot you’ll get roughly one keeper in every four generations. On a complex action beat, closer to one in twelve. A sixty-second edit with eighteen shots means somewhere between seventy and two hundred generations, so plan your credits before you plan your render.
Step 6: Rebuild the budget with real numbers
Subscription-included credits made Sora feel free. Pay-per-second APIs don’t. A typical replacement runs somewhere between ten cents and a dollar per second of generated footage, so that sixty-second piece at a 4:1 waste ratio can land anywhere from $30 to $300 in generation alone, before editing.
Two levers control that. First, generate at the lowest resolution that survives your delivery spec, then upscale. Second, cut the number of shots. Eight well-chosen shots with strong camera moves will outperform eighteen mediocre ones, and they’ll cost a third as much.
If the whole episode left you wondering whether AI video was ever going to be dependable enough to build on, the case for treating Sora’s shutdown as a reality check makes a fair argument, and it’s a useful read before you promise a client anything on a tight deadline.
Step 7: Insulate against the next shutdown
Assume every model you use today will be retired within two years, probably with less warning. The defence is boring and takes an afternoon. Keep prompts in plain text, outside any vendor’s interface. Write your deliverable specs down separately from your generation settings. Log which model produced which shot, so you know exactly what has to be re-rendered if one disappears. Once a month, run a single shot through a second model and compare. That twenty-minute test is the difference between a one-day migration and a lost fortnight.
One small mercy: search results for the name have gone strange. Type “Sora” into a search engine now and you’re as likely to get Bluetti’s Sora 500 solar panel as anything to do with video. Bookmark your actual tools, or your next research session will be about camping equipment.