OpenAI's deprecation table names a successor model for almost every retired API. For the Videos API and Sora 2, the replacement column is empty. That blank space is the most informative thing published about AI video this year, and it has a deadline attached: 24 September 2026.
OpenAI notified developers on 24 March 2026 that the Videos API and the Sora 2 model family — including sora-2, sora-2-pro and their snapshots — would be removed on 24 September 2026, and its deprecation table names no recommended replacement. The consumer Sora apps closed on 26 April 2026. It is a hard cutoff with no announced grace period, so export assets before evaluating alternatives.
What is being switched off, and when
OpenAI notified developers on 24 March 2026 that the Videos API and the Sora 2 model family would be deprecated and removed from the API on 24 September 2026. That covers sora-2, sora-2-pro and their dated snapshots. The consumer-facing Sora web and app experiences were discontinued earlier, on 26 April 2026, so the API is the last piece still running.
It is a hard cutoff rather than a rolling deprecation. There is no announced grace period, and OpenAI has advised users to export their Sora content. If your generated assets live only in OpenAI's storage, they are a download task with a deadline, not an archive.
Two weeks is enough time if you start this week. It is not enough time if you find out in October.
Why "no recommended replacement" is the detail that matters
OpenAI's deprecation table has a column naming the successor model for each retired one, and for almost every row it is filled in. For the Videos API and Sora 2 row it is empty.
That is not an oversight and it is not a small thing. It means OpenAI is not upgrading its video generation offering, it is leaving that market for now. Anyone whose content pipeline assumed "the video model we use will keep improving, like the text models did" has just had that assumption falsified in the clearest possible way.
The planning conclusion is worth more than the migration itself: in a content stack, the video model is the least stable component. Do not couple your production process to one vendor's API shape. Keep your prompt library, reference images, brand rules and shot lists in your own repository, in a form a human can read, and treat the model as a swappable back end. Teams that did this will spend two days on the migration. Teams that embedded prompts and parameters in application code, or that only have the outputs, will spend two weeks.
Where to go, by what you actually make
Almost every comparison of AI video tools ranks them overall. That ranking is not useful, because these models are not better and worse than one another in general — they are differently shaped, and the right answer depends entirely on the job.
| What you make | Where it goes | Why | Watch out for |
|---|---|---|---|
| Social cutdowns with dialogue or native audio | Google Veo (3.1 family) | Closest feature parity to Sora 2, with native audio generation rather than a separate pass | Terms and pricing differ substantially between the consumer app, AI Studio and Vertex AI. Establish which one you are on before you budget |
| Multi-shot sequences with a consistent character or product | Runway (Gen-4 family) | Reference image controls, character consistency across shots and an editor built around iteration rather than one-shot generation | Credit-based pricing makes per-project budgeting harder than a per-second rate |
| High-volume background plates and B-roll | Kling 3.0 | Lowest published per-second cost, strong native audio and lip-sync | Branding requirements and the governing-law terms. Read the licence before it goes into paid media |
| Anything that cannot leave your infrastructure | Self-hosted open-weight models | Complete control over inputs, outputs and retention | A real quality gap versus the frontier models, plus GPU cost and someone's time to run it |
| Hero product shots that must be exactly right | A camera | AI video still fails at exact product fidelity: logos, materials, text on packaging, hands. This has not changed in 2026 | Nothing — this is the correct answer more often than the tooling discourse suggests. What that actually costs is usually less than teams assume |
Cost per usable clip, not cost per second
Every comparison of these tools leads with the per-second price. Published list prices in September 2026 sit roughly here, and they change often enough that you should verify before committing a budget: Veo 3.1 Lite without audio around $0.05 per second, Kling 3.0 around $0.09 to $0.14, Runway Gen-4.5 around $0.15, and Veo 3.1 Standard with audio around $0.40.
Per-second price is the wrong unit, because you do not ship every generation. You ship the ones that work. The number that decides your bill is the keeper rate: how many generations you burn to get one shot you would actually put in front of a customer.
Cost per usable clip = (price per second × clip length) ÷ keeper rate. A keeper rate of 1-in-3 is realistic for an experienced operator with a loose creative brief. A keeper rate of 1-in-8 is realistic for a brand-exact brief with a specific product, a mandated colour and on-screen text.
Run that through an eight-second clip, the standard unit for social:
| Model | List $/second | One 8-second generation | Per usable clip at 1-in-3 | Per usable clip at 1-in-8 |
|---|---|---|---|---|
| Veo 3.1 Lite (no audio) | ~$0.05 | $0.40 | $1.20 | $3.20 |
| Kling 3.0 | ~$0.12 | $0.96 | $2.88 | $7.68 |
| Runway Gen-4.5 | ~$0.15 | $1.20 | $3.60 | $9.60 |
| Veo 3.1 Standard (with audio) | ~$0.40 | $3.20 | $9.60 | $25.60 |
The keeper rates are our planning assumptions from production work, offered as a model to fill with your own numbers rather than as measured data. Count your own for one project and the model becomes precise.
Now the conclusion, which runs directly against how these tools are usually compared. A thirty-second social cut needing six usable shots costs somewhere between about $7 and $154 in generation fees across that whole table. The same cut costs an operator four to six hours. At any realistic hourly rate, the human time exceeds the model fee by one to two orders of magnitude, in every row.
Which means choosing a tool on its per-second price is optimising the smallest line in the budget. The variables that actually move your cost are the keeper rate and how quickly your operator reaches a usable shot — both of which are functions of the tool's controls and your team's familiarity with them, not its rate card. A model that costs three times as much per second and halves your generation count is cheaper, and a model your editor already knows is cheaper still.
Before you commit, run the same brief through your two shortlisted tools: one specific product, one mandated colour, one line of on-screen text, ten generations each. Count keepers. That two-hour test tells you more about your actual cost than any published comparison, because it measures your brief and your operator rather than a generic prompt.
The rights question every comparison skips
Feature and price comparisons are everywhere. Licensing comparisons are not, and licensing is the dimension that determines whether an asset can legally carry a paid media budget. Six questions, in the order they cause problems:
| Question to ask the vendor | Why it matters for brand use |
|---|---|
| Do we own the output, or hold a licence to use it? | Determines whether you can sub-license it to a retail partner, include it in an asset handover, or assert anything against a copycat |
| Can you train on our inputs and outputs? | Your inputs include unreleased product photography and brand briefs. This is the clause that most often fails an internal review |
| Is there a mandatory watermark or attribution requirement? | A visible vendor mark disqualifies a tool for most paid media, and some licences require it on the paid tier |
| Which jurisdiction and which arbitration venue govern the terms? | Only matters once, and then it matters a great deal |
| Does the tier we are on permit commercial use at all? | Free and preview tiers frequently do not, and preview-tier restrictions have caught out teams who assumed the paid plan's terms applied |
| Does output carry machine-readable provenance metadata? | Needed for platform AI labels and for EU transparency obligations |
The specifics as they stand, with the caveat that these terms change frequently and you should read the current version rather than this paragraph: Runway's terms claim no ownership of outputs and place no restriction on commercial use, while still permitting the company to use inputs and outputs to improve its models. Kling grants a usage right rather than clear ownership, with a branding requirement and terms governed outside the EU and US. Google's commercial position differs depending on whether you are on a consumer plan, an AI Studio preview or Vertex AI, and preview terms have carried commercial-use restrictions that the generally available product does not.
One practical habit fixes most of this: save a dated PDF of the terms you relied on alongside the asset. When someone asks in eighteen months whether that campaign video was cleared, the answer is in the folder rather than in someone's memory of a page that has since been rewritten.
Separately, and regardless of vendor: synthetic media shipped into the EU after 2 August 2026 falls under the AI Act's transparency obligations, which is a visible label and machine-readable provenance, not a line in your terms of service. We covered the practical implementation in the EU AI Act transparency rules, and the platform-side disclosure rules are a separate layer again.
The two-week migration checklist
- Export everything, first. Before any tool evaluation. Deleted data does not come back after 24 September, and this is the only step with an unmovable deadline.
- Extract the prompt library into your own repository. Prompts, reference images, negative prompts, seeds, shot lists, brand rules. In plain text, in your own version control, readable by a human. This is what makes the next migration cheap.
- Run the keeper-rate test. Same brief, two shortlisted tools, ten generations each, count what you would actually ship.
- Check rights for the specific tier you will buy. Not the marketing page. The terms for the plan, saved and dated.
- Re-verify brand safety on the new model. Different models fail differently. Re-run whatever checks you had for hands, text rendering, logo fidelity and unintended likenesses.
- Update the disclosure and labelling step in your SOP. Provenance metadata differs between vendors, and your labelling process probably assumed Sora's.
- Rewrite the SOP itself to be vendor-neutral. "Generate the plate" rather than "open Sora". This is the step everyone skips and the reason the next shutdown costs as much as this one.
What this should change about planning AI video at all
A frontier lab launched a video product to enormous attention, shut down the consumer apps thirteen months later, and is now removing the API with no successor named. That is the clearest available evidence about how settled this category is, and it should inform how much of your content operation depends on it.
The evidence on effectiveness points the same way. Consumer trust in undisclosed AI-generated marketing fell through 2026 while user-generated content continued to outperform polished brand content on conversion, which is why the durable pattern has been AI behind the camera and humans in front of it: AI for plates, backgrounds, variants, localisation, rough cuts and versioning, humans for the faces, hands and claims. That division of labour survives a vendor shutdown, because it never depended on any particular vendor.
If Sora's removal costs you two weeks, the lesson is not which tool to buy next. It is to build the pipeline so the next one costs two days.
Frequently asked questions
When exactly does the Sora API shut down?
24 September 2026. OpenAI notified developers on 24 March 2026 of the deprecation and removal of the Videos API and the Sora 2 video generation model aliases and snapshots, covering sora-2, sora-2-pro and their dated versions. The Sora web and app experiences were discontinued earlier, on 26 April 2026. It is a hard cutoff rather than a rolling deprecation, with no announced grace period.
What is OpenAI's recommended replacement for Sora 2?
There isn't one. OpenAI's API deprecations page has a column naming the successor for each retired model, and it is filled in for almost every entry. For the Videos API and Sora 2 row it is empty, which means OpenAI is exiting video generation in the API rather than upgrading it. Teams must migrate to a third-party model rather than to another OpenAI product.
Which AI video tool should replace Sora for brand content?
It depends on the job, not on an overall ranking. Google Veo 3.1 is the closest feature parity for social cutdowns with native audio. Runway's Gen-4 family is the strongest for multi-shot sequences needing a consistent character or product, because of its reference image controls and editor. Kling 3.0 is the cheapest per second for high-volume B-roll and background plates. Anything that must stay on your own infrastructure means a self-hosted open-weight model and a real quality trade-off.
How much does AI video generation actually cost per clip?
Published list prices in September 2026 run roughly $0.05 per second for Veo 3.1 Lite without audio, $0.09 to $0.14 for Kling 3.0, about $0.15 for Runway Gen-4.5, and about $0.40 for Veo 3.1 Standard with audio. But the per-second price is the wrong unit, because you only ship the generations that work. Divide by your keeper rate: at one usable clip in eight, an eight-second Runway clip costs about $9.60 rather than $1.20.
Is the cheapest model per second the cheapest overall?
Usually not. A thirty-second social cut needing six usable shots costs somewhere between roughly $7 and $154 in generation fees across the main models, while costing an operator four to six hours. At any realistic hourly rate the human time exceeds the model fee by one to two orders of magnitude. A model that costs more per second but halves your generation count, or that your editor already knows, is cheaper in practice.
Can we use AI-generated video in paid advertising?
It depends on the vendor, the specific tier you are on and where the ad runs. Check whether you own the output or hold a licence, whether the vendor can train on your inputs, whether a watermark or attribution is required, and whether your plan permits commercial use at all — free and preview tiers frequently do not. Separately, synthetic media shipped into the EU after 2 August 2026 carries AI Act transparency obligations, meaning a visible label and machine-readable provenance. Save a dated copy of the terms you relied on alongside the asset.
The takeaway
Export your assets this week, because that is the only step with a deadline you cannot move. Then take the longer lesson: a frontier lab launched a video product to enormous attention and is removing the API thirteen months later with no successor named, which tells you the video model is the least stable component in a content stack. Keep prompts, reference images and brand rules in your own repository, choose the next tool on keeper rate rather than rate card, read the licence for the tier you are actually buying, and write the process so the model is swappable. If this migration costs you two weeks, the goal is for the next one to cost two days.
Sources & further reading
- OpenAI, “Deprecations” (API documentation)
- OpenAI Help Center, “What to know about the Sora discontinuation”
- Futurum Group, “OpenAI Sora Discontinuation: What the End of a Platform Means for Enterprise AI Strategy”
- Rangy, “AI Video Pricing Explained (2026): APIs vs Subscriptions”
- BuildMVPFast, “AI Video Generation API Pricing: Veo 3.1, Kling 3.0, Seedance, Runway Costs”
- LicenseOrg, “The Complete Guide to AI Commercial Use in 2026: Images, Video, Music, and Voice”