Short answer: there is no Veo 4 and no announced release date for one. More usefully, Veo is no longer Google’s recommended video model. Google’s own API documentation states that the Gemini API offers two video generation models, and tells developers to use Gemini Omni Flash as the default model for video generation, reserving Veo 3.1 for specific capabilities like scene extension, last-frame control, or integration with legacy pipelines. If you are waiting for a version number, the more important development already happened under a different name.
The Version Number Is the Wrong Thing to Track
Searches for Veo 4 assume a linear product line: Veo 3, then Veo 3.1, then Veo 4, each better than the last. That is not what happened.
What happened is that Google added a second video model from a different family and made it the recommendation. Veo 3.1 remains available and supported, and its role changed from “the video model” to “the model for particular jobs”.
The documentation phrasing is worth noticing. Alongside scene extension and last-frame control, Google lists integration with legacy pipelines as a reason to choose Veo. When a vendor describes one of its own models as the legacy-pipeline option, that is a clearer statement of direction than any roadmap.
What that means practically
If you were waiting for Veo 4 to start a project, stop waiting and evaluate Gemini Omni Flash instead. If you already use Veo 3.1 for extension or frame control, keep using it, because those are the capabilities it is now positioned for.
The Two Models, Side by Side
| Gemini Omni Flash | Veo 3.1 | |
|---|---|---|
| Google’s recommendation | Default for video generation | Specific capabilities and legacy pipelines |
| Strengths | Coherence, character consistency, factual accuracy | Native audio, extension, frame control |
| Inputs | Text, images, audio and video simultaneously | Text and image-based direction |
| Editing | Multi-turn conversational editing | Frame-specific generation |
| API surface | Interactions API for multi-turn refinement | generateContent API |
| Best for | Iterative work where you refine across turns | Precise control over start and end frames |
Two rows deserve expanding.
Multi-input reasoning
Google describes Gemini Omni Flash as supporting text, images, audio and video inputs simultaneously. That is a meaningfully different interface from text-plus-image prompting, because it lets the model reason across modalities rather than treating a reference image as a style hint.
Conversational editing
The documented ability to refine results across multiple turns, with examples including element replacement and perspective changes, addresses the biggest practical annoyance in generative video: getting something 90 percent right and having no way to fix the remaining 10 percent except regenerating and losing what worked.
Where Veo 3.1 Still Wins
Being specific about this matters, because the recommendation to default elsewhere is not an instruction to abandon it.
Native audio. Google describes Veo 3.1 as generating video with native audio, which for anything where sound and picture need to originate together is a genuine differentiator.
Video extension. Continuing an existing clip rather than generating a new one. If your workflow builds longer sequences from shorter pieces, this is the capability you need.
Frame-specific generation. First-frame and last-frame control, which is how you make a clip start and end where a storyboard requires.
Existing integrations. If you have a working pipeline, Google explicitly names legacy integration as a valid reason to stay.
The Vertex AI documentation lists the shared capability set across Google’s video models: text to video, first frame image to video, first and last frames to video, generating from references, extending videos and editing videos.
How to Tell a Real Release From a Rumoured One
This question is a useful case study, because the rumour material around it is unusually elaborate. Pages exist describing Veo 4’s pricing per second, its regional rollout, and one contributor’s experience of a colleague still receiving the older model by default.
None of that is impossible-sounding. It is exactly the kind of detail a real staged rollout would produce, which is what makes it convincing.
The check that actually works
Ignore the description and look for the artefact. A released model has three things that cannot be fabricated by an article:
- An identifier. A string you pass to an API, listed in the provider’s own documentation.
- A documentation page. Not a blog post, a reference page with parameters, limits and examples.
- A consistent date. One date, repeated everywhere, matching the provider’s own announcement.
Run those three checks against any model claim and the answer takes two minutes. Run them against Veo 4 and the API documentation names two models, neither of which is Veo 4.
Why staged-rollout claims are especially unreliable
“It is rolling out gradually, you may not have it yet” is unfalsifiable, and that is precisely why it appears so often in this genre. It explains any absence of evidence, and it lets a page describe a product nobody can check.
Genuine staged rollouts still have documentation, still have an identifier, and are still described by the vendor. When a rollout claim comes with none of those, the simplest explanation is usually correct.
What to Know Before You Trust a Release Date
Search for Veo 4 and you will find pages describing its features, its pricing per second, and even reviews of a staged regional rollout. Apply these checks.
Does an official Google source name it? The Gemini API documentation names two video models, and Veo 4 is not one of them.
Can you call it? A released model has an identifier you pass to an API. No identifier, no model.
Do the claims agree? Pages announcing Veo 4 place its release in different months and describe different capabilities. Genuine launches do not produce that pattern.
Is there documentation? Google publishes model documentation on release. An announcement with no docs page is not a release.
Check the page date against the tense. Articles dated before a supposed launch that describe it in past tense are recycled speculation.
Why Vendors Keep Doing This
It is worth understanding the pattern, because Veo is not an isolated case.
Model families get renamed, merged and repositioned for organisational and strategic reasons rather than technical ones. A new team ships a new family. A capability moves from a specialist product into a general assistant. A model line that was the flagship becomes the compatibility option.
For anyone building on these tools, that means version numbers are a poor planning unit. What matters is which identifier you can call today, what it does well, and how much work it would be to swap.
A Decision Framework for Choosing a Video Model
- Start from Google’s own recommendation. If you have no specific requirement, use the default rather than the familiar name.
- List your hard requirements first. Native audio, clip length, frame control, extension. These eliminate options faster than quality comparisons.
- Test on your own difficult material. Faces, motion, on-screen text and consistency across shots are where video models diverge most. Showcase reels tell you nothing about your content.
- Check the editing model. Whether you can refine an output or must regenerate is the biggest practical difference in day-to-day use, and it rarely appears in comparisons.
- Price per second, not per generation. Video costs scale with duration in ways text does not, and a per-clip figure hides that.
- Assume you will migrate. Keep prompts, reference assets and style notes outside the vendor’s interface.
Common mistakes right now
- Waiting for a version number instead of evaluating what is available now.
- Sticking with a familiar model after the vendor has repositioned it.
- Judging models on demo reels rather than on your own hardest footage.
- Ignoring the editing workflow, which determines how many attempts a result actually costs.
- Treating a confident third-party review of an unreleased model as evidence it exists.
What Google Has Actually Been Shipping
For context on the pace, the Gemini app’s own release notes record a steady stream of updates through 2026: a new Flash model generation, an autonomous agent extended to desktop, voice dictation with screen context on macOS, and a dedicated student hub with study notebooks. That is a company shipping frequently across many surfaces rather than saving everything for numbered flagship releases.
Video generation fits that pattern. The capability improved, the recommended model changed, and no version number was involved. Anyone tracking releases by major version was looking in the wrong place for most of the year, which is the general lesson worth taking from this specific question.
Building the Skill That Survives Renaming
Video model expertise dates quickly. What does not is the craft: describing motion rather than only a scene, iterating toward a result rather than regenerating hopefully, knowing which failures are prompt problems and which are model limits, and judging whether an output is usable rather than merely impressive.
Those transfer between every model in this category, and given how often the category renames itself, they are the part of your knowledge with the longest useful life. Learning them in a structured sequence, rather than as a series of product tutorials, is what makes the next repositioning a footnote instead of a relearning exercise. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
FAQ
Is there a Veo 4 release date?
Is Veo discontinued?
Which Google video model should I use?
Does Veo 3.1 still get updates?
Why do sites claim Veo 4 exists?
What to Do Instead of Waiting
Take the last video task you wanted to automate and run it through Google’s current default model rather than searching for the next version number. Then run the same brief through the model you were already using. Comparing two real outputs on your own material will settle the question in twenty minutes, which is twenty minutes better spent than any amount of waiting for a release that has not been announced and may never carry the name you are searching for.