For AI video, the most important changes between August and early November 2026 were not only new models but also shutdowns. OpenAI’s Sora API reached its listed discontinuation date, Google retired preview model IDs for Veo 3.1, and Kling published its 4.0 capabilities. If you built a workflow on any of those, the quarter may have required action rather than just curiosity.
This article lists the dated changes we could verify, explains what each means for someone choosing or maintaining a video workflow, and says plainly where our evidence is thin. We rely heavily on one release tracker, so we flag that throughout.
How we compiled this: we did not run hands-on tests. This is based on a single dated release tracker and public reporting, checked on 11 November 2026. We could not open most video vendors’ own news pages, so vendor details below are second-hand unless stated. Always confirm on the vendor’s site.
Key takeaways
- A release tracker lists 24 September 2026 as OpenAI’s shutdown date for the Sora API (Videos API, sora-2, sora-2-pro), and says the Sora web and app experience ended on 26 April 2026.
- The same tracker lists 22 October 2026 as Google’s shutdown date for the Veo 3.1 Preview model IDs, pointing new direct Gemini video work to Gemini Omni 1.2 Flash.
- New or updated models in the window, per the tracker: MiniMax H3 open weights (3 August), LTX-2.5 (11 August), Gemini Omni 1.1 Flash (27 August) and Kling 4.0 (15 October).
- The tracker states that listed capabilities are the providers’ own, not a quality benchmark, and we have not tested any of these tools.
- Practical lesson: model identifiers and APIs get retired, so check deprecation dates before you build on a video model.
Summary
Our one dated source for this category is a release tracker published by Magic Hour, a company that sells its own AI video product. It was first published on 8 March 2026 and cites verification dates up to 10 October 2026. Magic Hour says its listings describe provider capabilities rather than its own quality benchmark. Because the tracker comes from a vendor in the same market, we use it for dates and provider-documented facts only, and we do not repeat any quality claims.
From it, three kinds of change stand out: retirements of older access routes, open or developer-facing model releases, and one provider’s updated capability set.
Shutdowns and retirements
OpenAI Sora API, 24 September 2026
The tracker lists 24 September 2026 as OpenAI’s shutdown date for the Sora API, covering the Videos API, sora-2 and sora-2-pro, and notes that the date has passed. It also says the Sora web and app experience ended on 26 April 2026. We could not open an OpenAI page to confirm these dates, so check OpenAI’s own documentation if you used Sora through an app or the API.
Google Veo 3.1 Preview IDs, 22 October 2026
The tracker lists 22 October 2026 as Google’s shutdown date for the Veo 3.1 Preview model IDs and for an older Omni Flash preview ID. It says Google points new direct Gemini video work to Gemini Omni 1.2 Flash or to available Cloud GA models. It also notes that requests to Gemini Omni 1.1 are set to route to 1.2 and bill at 1.2 rates from 3 December 2026, a date outside our window but relevant if you use 1.1 today. The tracker cites Google’s deprecations page, updated 9 October, but we did not open that page ourselves.
What it means: if your scripts or no-code tools call preview model IDs, they may have stopped working or may soon. Check the model name your tool uses and ask the vendor which version is behind it. This is a good moment to read our note on chatbots versus agents too, since agents that call video tools inherit the same risk.
New models and releases
| Date (2026) | Name | What the tracker says |
|---|---|---|
| 3 August | MiniMax H3 open weights | MiniMax released H3 model weights; its hosted Context-IR and Regenerate-2K stages remain separate services |
| 11 August | LTX-2.5 | Lightricks’ official LTX-2 release (v1.2.0) added LTX-2.5 checkpoint support; docs list Fast and Pro variants with output up to 4K |
| 27 August | Gemini Omni 1.1 Flash | Google released the stable gemini-omni-1.1-flash API model with extension, first and last frame interpolation, and resolution control |
| 15 October | Kling 4.0 | Kling documents native 3 to 30 second generation, up to 15 combined reference assets and up to 10 keyframe inputs |
The tracker says the Kling figures are provider capabilities, not a Magic Hour quality benchmark. The same applies to our reading: these are documented specifications, not evidence of how good the output is.
Open weights and local use
Two of the four releases, MiniMax H3 and LTX-2.5, are described in terms of weights or checkpoints. That matters for teams that want to run video models on their own infrastructure, since it changes who holds the data. It also means setup, hardware and licence terms become your responsibility. The tracker does not give licence details, so read each project’s licence before using the output commercially.
Google: Gemini Omni
Google appears twice in the tracker. On 27 August it released the stable gemini-omni-1.1-flash API model, and by October it recommends Gemini Omni 1.2 Flash for new direct Gemini work. Separately, a May 2026 report on Google I/O, from Android Authority, said Gemini Omni was rolling out that day to Google AI Plus, Pro and Ultra subscribers worldwide. That is before our window, but it shows Omni is both an app feature and an API model family. We did not find a Google page in this research describing the differences between Omni 1.1 and 1.2.
What it means: if you use Gemini for video through the app, you probably will not see model IDs. If you use the API, you will, and the retirement of preview IDs makes it worth pinning to a stable version and watching deprecation notices.
Timeline at a glance
| Date (2026) | Provider | Event | Type |
|---|---|---|---|
| 3 August | MiniMax | H3 open weights released | Release |
| 11 August | Lightricks | LTX-2.5 support in LTX-2 v1.2.0 | Release |
| 27 August | Gemini Omni 1.1 Flash stable API model | Release | |
| 24 September | OpenAI | Sora API shutdown date | Retirement |
| 15 October | Kling | Kling 4.0 capabilities documented | Release |
| 22 October | Veo 3.1 Preview IDs shut down | Retirement | |
| 3 December | Omni 1.1 requests to route to 1.2 (upcoming) | Scheduled change |
What it means for tool buyers
- Check the lifespan of what you build on. Two of the six dated events above are retirements. Before you build a workflow, ask whether the model is a preview, whether a deprecation date exists, and what the vendor’s notice period is.
- Separate documented specs from quality. Clip length, reference counts and keyframes are specifications. They do not tell you whether the video looks good. Our guide to comparing AI tools side by side shows how to judge quality using your own prompts.
- Read the pricing unit. Video tools charge by seconds, credits or tokens, and Google says Omni 1.1 will bill at 1.2 rates from December. See how to read an AI tool’s pricing page.
- Treat open weights as a different decision. Running a model yourself changes cost, privacy and responsibility. Read how to protect company data.
- Plan prompts and scripts. If you produce video from text, our guides on AI video prompts and video scripts cover the writing side, which is independent of which model you use.
A note on sourcing
This is the thinnest article in our quarterly series, and we would rather say so than pad it. Our searches for video news returned many general “best of” pages, and the few we opened either predated the window or did not give dated releases. We could not open the news pages of Runway, Luma, Google DeepMind or OpenAI. We therefore have no verified information in this article about, for example, Runway or Luma releases in the period, and the absence of an item here should not be read as meaning nothing happened.
We also did not read the original provider pages for the items in the tracker. If a date matters to your project, such as a shutdown that could break a workflow, verify it with the provider.
What to watch next
These are open questions from the sources, not predictions.
- Whether a replacement route for OpenAI’s retired Sora API is announced, which the tracker does not mention.
- How the 3 December routing change from Gemini Omni 1.1 to 1.2 affects pricing for existing users.
- Whether independent evaluations of Kling 4.0, LTX-2.5 and MiniMax H3 appear, so documented specifications can be compared with actual output.
Frequently asked questions
Is Sora still available?
According to the tracker, the Sora web and app experience ended on 26 April 2026 and the Sora API had a listed shutdown date of 24 September 2026. Confirm with OpenAI.
What happened to Veo 3.1 Preview?
The tracker lists 22 October 2026 as Google’s shutdown date for the Veo 3.1 Preview model IDs and points new work to Gemini Omni 1.2 Flash or Cloud GA models.
What is new in Kling 4.0?
The tracker lists documented capabilities: native 3 to 30 second generation, up to 15 combined reference assets and up to 10 keyframe inputs. These are provider specifications, not test results.
Are any of these models open?
The tracker describes MiniMax H3 as open weights, and LTX-2.5 as a checkpoint in Lightricks’ official release. Check licence terms before commercial use.
Which AI video tool is best?
We cannot say, because we have not tested any of them. Use your own prompts and compare outputs.
Why is this article shorter than others?
Because we could verify only a small number of dated events. We chose a shorter, accurate piece over padding.
Conclusion
For AI video, the verified story of the quarter is as much about retirement as release: older access routes were closed, and newer models arrived with documented specifications that remain to be judged. Before you commit, read how to choose an AI tool, and run a trial with our pre-purchase checklist.
Sources
- Magic Hour, AI video model release tracker 2026, accessed 11 November 2026
- Frameo, AI video model updates (published 13 May 2026), accessed 11 November 2026
- Android Authority, Gemini updates at Google I/O 2026, accessed 11 November 2026
