The most useful thing that happened to AI video this year was not a better model. It was a reminder that the tools are not permanent.
OpenAI announced in March 2026 that it was shutting down Sora — the app first, with timelines for the app and the API to follow, and the much-publicised Disney licensing arrangement collapsing with it. Sora 2 was, by most accounts, one of the strongest video models available. That did not save it, because a model is not a product and a product is not a business. By late 2026, reports had access to Sora 2 and the video API winding down entirely, exactly on the schedule the company had signalled.
For anyone building a production pipeline, the lesson is more valuable than any new capability: never let your delivery depend on a single vendor's roadmap.
What the landscape looks like now
Three shifts matter for production work, and none of them are about raw quality.
Google's Veo line became the default. Veo 3.1 offers four, six or eight second clips, 720p and 1080p input with 4K output available, up to four outputs per prompt, native sound generation, and 9:16 or 16:9 — which is to say, exactly the shapes a social production pipeline needs. It also carries C2PA Content Credentials, meaning the generated file records that it was generated. That last detail is quietly important for any agency working with regulated clients.
Kling moved from "cheap alternative" to "the iteration tool". Kling 4.0 is built as a native multimodal model with more stable complex motion and better subject consistency across a sequence, and the Flash tier prioritises speed at 720p. For a production team the practical difference is simple: Veo for the shot that has to be right, Kling for the sixteen takes you need before you find it.
Open-weight models created a cost floor. Fine-tunable open models mean a studio with a consistent style can train and run its own, which caps what the closed vendors can charge for volume work. Most client projects do not need this. The ones with a distinctive, repeated look do.
| Capability | A year ago | Now | What it changes |
|---|---|---|---|
| Clip length | 4-8s, awkward joins | 8s clean, longer sequences in tools | Fewer stitches, fewer artefact seams |
| Native audio | Novel | Standard | Dialogue and sound design in one pass |
| Character consistency | Per-prompt lottery | Improving across a sequence | Repeat presenters become viable |
| Provenance | Optional | C2PA on major models | Agencies can answer "was this AI?" |
| Model availability | Assumed stable | Already reshuffled once | Vendor risk is a planning item |
| Prompt languages | English-first | Still largely English | Non-English briefs need translating first |
The detail most teams missed: retirement dates
Veo 3.1 went generally available in November 2025, and its documentation already lists a retirement date of May 2027 or later for support on Google Cloud, with a note that the model may remain accessible through other routes after that.
Read that carefully. The best-in-class model of late 2025 has a published expiry inside eighteen months. That is not a criticism of Google — it is normal platform lifecycle — but it does mean that any workflow with a model checkpoint baked into its delivery promise has a dated problem in it.
The practical response is not to avoid major vendors. It is to keep three things portable:
- The client brief, so a change of model does not change the offer.
- The footage library, so real product footage keeps working when generated scenes are regenerated.
- The review process, so quality assurance is a human step rather than a model setting.
Those three transfer overnight. Studios that lost nothing in March were the ones whose process was portable.
The things that did not change
This is the part of the review that matters more than the release notes. Three years of capability gains have not moved these:
The hook still decides everything. A better model produces a more beautiful ad that still gets scrolled past if the first two seconds do not name the viewer's problem. Everything in the Dubai hook bank would work on footage from 2023.
Product accuracy is still the failure mode. Generated footage still invents buttons on interfaces, changes label text, and gets the way a product opens wrong. Real footage of the product remains non-negotiable, and it is the reason a generated product shot is a legal exposure rather than a creative choice.
Review is still the bottleneck. A pipeline that generates forty clips in an hour still waits three days for a human to approve them, and that wait is what the client experiences. Capacity is not the constraint; judgement is.
Cost per finished video is still mostly labour. The model costs cents. The concept, the edit, the captions, the native review, the revision rounds and the QC are what you are paying for at $70 and up. Anyone selling "AI video" as a cheap product is selling the model. Anyone running a production business is selling the process.
What the model releases did change
Four genuine improvements worth using, if you know where to point them:
1. Eight seconds in one take is enough for a complete social beat — hook, action, result — without a visible join. That removed a whole class of edits from the pipeline.
2. Native audio means a two-second sound design pass is now standard rather than a separate vendor. It also means a generator can produce a voice that is fluent and wrong, which is why native review became more important, not less.
3. Better subject consistency makes a repeat presenter possible. This matters for serialised formats — a weekly explainer, a changelog series — where the same face across ten clips is now achievable without a shoot.
4. Provenance credentials turn "was this AI?" from a debate into a metadata check. Regulated industries are starting to ask, and having an answer is cheaper than having an argument.
None of these change what a good ad is. They change how fast you can make sixteen attempts at one.
What this means if you are buying, not making
Four questions worth asking any production partner, all of them downstream of the last 90 days:
1. Which models do you use, and what happens when one is retired? A partner with a single-vendor pipeline and no fallback has a business-continuity problem they have not noticed.
2. Can you show provenance on delivered files? With C2PA credentials standard on major models, this is now an answerable question.
3. How much of the product on screen is real footage? If the answer is "all of it is generated", the product is being misrepresented, which is an exposure for the brand that publishes it.
4. Who reviews the language? Model prompt handling is still overwhelmingly English-first. Arabic, German, French, Hebrew, Polish and Greek review is a human cost, and a partner who does not mention it is not doing it.
The 90-day forecast that actually matters
The pattern of the last year suggests the next 90 days will bring better character consistency, longer usable sequences and more provenance tooling. It also suggests that at least one model you use today will change materially — be retired, re-priced, or merged into something else.
So the practical resolution for a production team is unglamorous: keep the client brief, the footage library and the review process independent of any single model. Nothing else about the pipeline needs to be defended, because everything else is replaceable by design.
That is the design principle behind how to create AI videos, and the reason our own pricing is quoted per finished video rather than per generation credit. The model is an input. The delivery is the product.
How to evaluate the next release
When the next model lands — and one will, within the quarter — test it against your own footage rather than against a demo reel. Three checks, in this order: does it hold your product's geometry across an eight-second take, does it keep the on-screen text you supply instead of inventing it, and does it produce a still frame good enough to be the thumbnail. If any of the three fails, the model is not relevant to your pipeline yet, however impressive the launch video was.
What to do about it this month
Audit your pipeline for single points of failure. List every vendor in your delivery chain, and for each one write down what you would do if it disappeared in thirty days. Two entries will have no answer, and those two are your project for December.
What to keep in the toolbox
If you are running production rather than writing about it, five settings and habits are worth keeping regardless of which model you use this quarter:
- One reference image per product, kept forever. Consistency between clips comes from the same reference, not from the same prompt. Store it beside the footage, not in a chat history.
- Prompt in English even for Arabic or Polish output. Prompt handling is still overwhelmingly English-first; write the direction in English and localise the output, not the prompt.
- Generate the setting, record the product. The rule has survived three model generations and every product-launch cycle in that time.
- Keep a screenshot of the vendor's status page when a delivery slips. Clients accept a platform incident; they do not accept a shrug, and the difference is evidence.
- Price per finished video, never per generation credit. A per-credit price makes the client's bill a function of your retries, which is the fastest way to lose the account after a bad week.
None of these are exciting, and none will appear in a launch post. They are the parts of the pipeline that determine whether the business survives a model change, which is the only variable this year actually tested.
The White-Label Buyer’s Checklist
Twelve questions to ask any production partner before a client’s calendar depends on them — scope, rights, confidentiality and the one answer that should end the conversation.
or download it right now — no email needed.
One email, the PDF as a link. Reply and a person answers.
Your pipeline should outlive the model running it. Book a 20-minute call and we will map which parts of your video production survive a vendor change — and which do not.
Drop your take 0
// no hot takes yet — start the thread 👇