The most consequential detail in OpenAI’s Sora deprecation notice is the replacement column. The Videos API and the listed Sora 2 models are scheduled to shut down on September 24, 2026. No replacement is listed.
For an application that depends on those models, that is a migration deadline. Meanwhile, Runway’s Solaris announcement raises a different development question: what changes when a video model generates an interface that responds to you?
Both stories matter to people building with AI video. They require different kinds of work.
First, identify which Sora dependency you have
OpenAI’s Help Center separates the consumer product from the developer service. The Sora web and app experiences ended on April 26. The API shutdown is still scheduled for September 24.
Search your application for sora-2, sora-2-pro, dated model snapshots, and Videos API calls. Include scheduled jobs, internal tools, and configuration outside the main application repository. An old experiment can become a production dependency without acquiring much documentation along the way.
For each match, record what a successful result actually means. Is it a video file saved to your storage? A URL handed to another service? An email sent when generation finishes? A callback that advances a larger job?
That last part matters. Finding an alternative model that produces an attractive clip does not prove that the surrounding job can finish.
Define the migration’s acceptance criteria
A useful migration review follows a complete job from input to delivery:
- Input: Does the candidate accept the reference material, aspect ratio, duration, and audio requirements your feature needs?
- Job handling: What happens during a long wait, an explicit failure, or an interrupted connection? Can your application tell a failed job from a job that is still running?
- Output: Can you retrieve the result and pass it through the existing storage, editing, and delivery steps?
- Cost: What does an accepted result cost after rejected takes and repeated attempts, rather than just one successful request?
For programmatic work, reAPI’s unified AI model API is an example of access to image and video models through a common integration. That changes where you connect; each candidate still needs to pass your acceptance criteria. A common API does not make two models behave identically.
Use a small set of representative jobs for the comparison. Keep the briefs and acceptance criteria consistent, but adapt the requests to the target model’s documented inputs. Copying one provider’s payload into another provider’s endpoint is not a meaningful compatibility test.
Make the fallback concrete, too. After the shutdown, switching a feature flag back to Sora will not restore the retired service. Your fallback may need to be an already validated provider, a manual workflow, or temporarily disabling a feature with a clear message.
Solaris changes what you ask the system to produce
Runway introduced Solaris on August 31 as its first Interface World Model. Its announcement describes a language model directing behavior while a video model generates responsive frames, conditioned on interactions such as clicks and drags.
That suggests a different way to evaluate the output:
| Question | A generated video clip | An interactive generated interface |
|---|---|---|
| What is delivered? | A finished sequence you can review | A session that changes as someone acts |
| What needs checking? | The chosen take and exported file | The response to each action, including unexpected ones |
| What makes a failure costly? | A rejected take or missed delivery | A user acting on an incorrect or inconsistent display |
This is a practical distinction between product forms, rather than a benchmark comparison. A good clip can be accepted and stored. An interface must keep working after the next input, including an input its designer did not expect.
Solaris is currently presented with an early-access request, and Runway identifies unresolved text, grounding, long-session, and accessibility challenges. Its announcement gives developers a research direction to investigate, not a documented Sora migration path.
A rendered confirmation is a new testing problem
Consider a hypothetical store with a generated interface. A customer drags two items into a basket, removes one, and changes the delivery address.
The engineering questions extend beyond whether those interactions look natural. Did the order state change? Does the displayed total match it? Can a keyboard user complete the same task? If the session reconnects, which version of the basket is authoritative?
These are requirements I would bring to an evaluation, not claims about a shipping Solaris checkout system. A visual demonstration cannot answer them on its own.
For a prototype, I would keep order records and confirmations explicit and testable while experimenting with how the interaction is displayed. The experiment then has a clear question: can generated interaction make the task easier without making its result ambiguous?
Keep creative exploration connected to the delivery workflow
Not every Sora dependency is an API integration. A creator may simply need a repeatable way to turn a visual brief into a usable shot. A browser workspace such as ClipDance’s AI video generator addresses that authoring task through text-to-video and image-to-video tools.
The acceptance test there is the finished piece: can the selected take be edited, combined with sound, and delivered in the form the project needs? That question can be answered now. It does not require waiting for generated interfaces to mature.
Handle the archive separately. OpenAI recommends exporting Sora content promptly and says users will be emailed if a final export window is offered. Do not assume the API deadline also defines the last opportunity to retrieve consumer content.
Before September 24, the useful deliverable is a working route from brief to saved result, with an owner and a tested fallback. Solaris can earn a place in a separate experiment where the goal is interaction itself.
If your video provider disappeared tomorrow, which part of that route would you have the most trouble replacing?
Top comments (1)
The empty replacement column beside the September 24 Videos API shutdown is the key signal: teams need to migrate the whole job contract, not just swap a model name. I'd inventory callbacks, stored outputs, retries, and delivery handoffs alongside
sora-2references, then run representative briefs against a validated fallback and measure cost per accepted result. Solaris makes the separation even sharper-once clicks and drags change generated frames, authoritative basket state, reconnect behavior, keyboard access, and rendered confirmation belong to the product architecture, because visual plausibility is no longer evidence that the.