Veo 3.1 is Google's current video model. There is no Veo 4. Google's own developer documentation, updated 9 September 2026, describes Veo 3.1 as a model for generating 8-second videos at 720p, 1080p or 4K with natively generated audio. Anything quoting Veo 4 specifications is describing something that was never released.
This is a short post about a small error, and the reason it's worth writing is that the error is instructive. Nobody invented Veo 4 maliciously. They inferred it.
What is Google's current video model?
Veo 3.1, and Google's own documentation is unambiguous about what it does.
Per the Gemini API docs, it generates 8-second videos at 720p, 1080p or 4K, with natively generated audio. Three capabilities are named specifically: portrait output, so you choose between 16:9 landscape and 9:16 portrait; video extension, letting you extend videos previously generated with Veo; and frame-specific generation, where you specify the first and last frames.
There's a Veo 3.1 Fast variant alongside it, positioned for speed and business use — programmatic ad generation, rapid A/B testing of creative concepts, high-volume social content.
Google Cloud maintains separate model pages for Veo 3 and Veo 3.1. There is no page for a Veo 4.
Eight seconds, three resolution tiers, native audio. That's the model. There is no next one.
Worth noting how well that matches what Hexcoded's picker offers, because the two agreeing is the point of this post. The picker shows Veo 3.1 locked to 8 seconds in Create mode, with 720, 1080 and 4K, and two aspect ratios. Extend appends roughly seven seconds per run for up to around twenty runs — which is what Google means by "extendable past a minute."
Why do so many posts cite a Veo 4?
Three reasons, and none of them is deliberate invention.
Version inference. A model line that went 2, 3, then 3.1 looks like it's due a 4. Writers extrapolate the pattern and then describe what a 4 would plausibly do, which reads exactly like reporting.
Announcement confusion. Vendors announce capabilities that don't ship, or ship under a different name than the one previewed. A capability announced and then folded into a different product gets attributed to a version number that never existed.
Citation cascade. One post asserts it, the next cites the first, and after five iterations the claim has five sources and no origin. This is the mechanism that makes model misinformation durable.
A claim with five sources and no origin is still a claim with no origin.
Which is the practical lesson, and it generalises well beyond Google. Version numbers are the easiest thing in this category to get wrong and the easiest to check.
How do you check a model's actual version?
Go to the vendor, and note the date.
Google publishes Veo's capabilities on its own developer documentation, and every page carries a "last updated" line. The one describing Veo 3.1 was last updated on 9 September 2026. That's the authoritative source, and it's freely available — which makes citing a comparison post a choice rather than a necessity.
There's a second check worth running alongside it: open the picker you'll actually generate in. Vendor documentation tells you what a model can do; your own interface tells you what's available to you on your plan. Those aren't the same question, and 4K is a good example — Google offers it on Veo 3.1, and access to it may sit above your plan's tier.
- The version number, on the vendor's own model page rather than a review site
- The "last updated" date on that page
- Whether a capability is described as available or planned
- The resolution ceiling, which is announced optimistically more often than any other spec
- Whether the figure is per generation or reachable through chaining
- What your own plan actually permits, separately from what the model supports
Is this actually a problem worth caring about?
For casual reading, no. For a commercial delivery, yes.
A wrong version number is harmless until it's attached to a spec you planned around. Someone who promised a client 4K because a comparison post said a model delivered it has a schedule problem, not a trivia problem.
There's a live example in the other direction. Seedance 2.5 was announced with native 4K; the picker offers 480p, 720p and 1080p. That 4K figure is still repeated across comparison posts, and the announcement outranks the interface in search results.
So the failure runs both ways. A model can be credited with a version that doesn't exist, and a real model can be credited with a capability it never shipped.
What should you do with this post?
Distrust it, appropriately.
This post is a secondary source too. It's current as of the publication date, and the whole category moves monthly. If you're relying on a version number for a delivery, the point isn't that we told you Veo 3.1 — it's that Google's documentation says so, it says when it last said so, and you can read it yourself in thirty seconds.
Model versions and capabilities described here are current as of the publication date and change frequently. This post is a secondary source. Verify version numbers and specifications against vendor documentation and your own generation interface before relying on them.
- Veo 3.1 is Google's current video model. There is no Veo 4, and Google's own docs confirm it
- Google describes Veo 3.1 as 8-second videos at 720p, 1080p or 4K with native audio
- It offers portrait and landscape, video extension, and first-and-last-frame generation
- Any comparison quoting Veo 4 specifications is describing a model that was never released
- Version inference, announcement confusion and citation cascades produce most model misinformation. None requires bad faith
- Check the vendor's own page, and check its "last updated" date
- Then check your picker. What a model supports and what your plan permits are different questions
- The failure runs both ways. Seedance 2.5's announced 4K never shipped, and it's still quoted
No. Veo 3.1 is Google's current video model, and Google's own developer documentation confirms it. Comparison posts quoting Veo 4 specifications are describing something that was never released, usually by extrapolating from the version-numbering pattern.
Veo 3.1. Google's documentation describes it as generating 8-second videos at 720p, 1080p or 4K with natively generated audio, with portrait and landscape output, video extension, and the ability to specify first and last frames.
Eight seconds per generation. Longer runtimes come from chaining through Extend, which appends further footage to a clip previously generated with Veo — so "eight seconds" and "past a minute" are both true and describe different things.
Three mechanisms, none deliberate. Writers extrapolate from a version-numbering pattern; announced capabilities that never shipped get attributed to a version that doesn't exist; and one post citing another produces a claim with several sources and no origin.
The vendor's own model documentation, and the "last updated" date on it. Then your own generation interface, separately — vendor documentation tells you what a model can do, and your picker tells you what's available on your plan.
Yes, and in the other direction. Seedance 2.5 was announced with native 4K and shipped without it — the picker offers 480p, 720p and 1080p. So a model can be credited with a version that doesn't exist, or with a capability it never shipped.
Current specs, in front of you
Creative Studio runs 30+ models on one credit balance, with duration, resolution, aspect ratio and reference ceiling visible per model before you generate. The interface is the spec sheet.
Open Creative StudioMore on model capability, access and rights in Models.