Google Veo

Our verdict on Google Veo

Google Veo is the video model we have built the most around. When Veo 3 came out in mid-2025, Jordan called it a step-function improvement over Sora, and it pushed us to build our own team video studio. Veo 3.1 Fast later won our side-by-side test against Sora 2, but every version we used was slow or costly enough that you have to plan generations instead of firing them off freely.

What we built with Veo

• Ryz Labs Video Studio (June 2025): an internal web app built in Replit on Replicate that gives the whole team Veo 3, Veo 2, Seedance and image models in one place. The first version shipped in a few days. One of our video editors generated 20 to 30 clips in about an hour and had a capybara commercial for Offsiteio done within a day.
• Trick-or-treat AI costume builder (October 2025): a webcam app that restyles your photo with Nano Banana, then uses that image as the first frame of a Veo 3.1 Fast "Happy Halloween" video. The team used it heavily in its first 48 hours, and one teammate picked her real costume from it.
• AI compliance training and certification platform (November 2025): every scenario video in the exams was generated with Veo 3.1 Fast, on a Supabase back end. Sam estimated 60 to 70% of the video content was right on the first pass.

How we use Veo

1. Call it through an API, not a seat license. Our studio uses API keys as a consumption layer, so everyone gets Veo and other top models on a pay-per-use basis. For each model, Jordan copied the sample API call and JSON schema from Replicate and asked the Replit agent to add it (Episode 1).
2. Enhance the prompt first. "Enhance with AI" was the first feature we added after launch. It sends your short prompt and the target model to ChatGPT, which turns "dog skateboarding in Venice, Italy" into a detailed scene description (Episode 1, at 09:55).
3. Test on a cheaper model. At 12:39 in Episode 1, Jordan ran the same prompt on Veo 2 instead of waiting ten minutes for Veo 3. Save prompts with a copy button so you can rerun them anywhere.
4. Budget for retakes. A 30-second commercial needs at least four eight-second clips, and Jordan warned you could easily generate 10 and spend $60 to $100 in minutes.
5. Start from an image when the subject matters. The costume builder feeds the Nano Banana image into Veo 3.1 Fast as the source image, with a short Halloween prompt.
6. Edit the clips together like footage. Our editor stitched the capybara clips, layered a voiceover and shipped. Sam says we also use Veo 3 for b-roll to supplement original content (Episode 18, at 08:34).

It's about $6 per eight second generation, and the generations are not perfect.

— Jordan Metzner, Episode 1

Where Veo falls short

• Speed. Veo 3 sometimes took up to ten minutes for an eight-second clip in June 2025. Jordan said price mostly tracks generation time.
• Cost for a startup. Compared with a day of filming, Veo 3 is much cheaper. But Jordan said a startup can easily spend hundreds if not thousands of dollars if it isn't paying attention.
• Dialogue. In the costume app, the spoken lines were often mumbled non-words, which reminded Jordan of the garbled text in early AI images (Episode 18, at 10:31).
• First-pass accuracy. The compliance videos needed a few passes. Only 60 to 70% was right off the bat.
• Not TV-ready at scale. In September 2025, Jordan said none of the video models were good enough to make production-quality television at scale yet.

And, yes, it did require a few passes to get the content right. I would say 60 to 70% of the content was right right off the bat.

— Sam Nadler, Episode 20

Google Veo compared

• Sora vs Google Veo: Veo won on quality and physics in every comparison we ran; Sora 2 wins on putting your own face in the video.
• Seedance vs Google Veo: Seedance renders in moments; Veo 3 could take ten minutes and cost much more.

The October 2025 test is the one we would point to. Jordan ran the same two-capybara podcast prompt through his Sora 2 API video generator and through Veo 3.1 Fast in our studio. Both hosts judged the Veo clip better.

Google yesterday released Veo 3.1, and using our standard Ryz Labs video generator, Veo 3.1 fast is is pretty pretty good and pretty inexpensive.

— Jordan Metzner, Episode 17

Episodes featuring Google Veo

• Episode 1: why Veo 3 made us build a video studio, plus cost and speed numbers.
• Episode 10: Jordan says nothing has beaten Veo 3 Fast yet.
• Episode 15: first day with Sora 2, compared with our Veo 3 experience.
• Episode 17: Veo 3.1 Fast vs Sora 2 on the same prompt.
• Episode 18: the Nano Banana plus Veo 3.1 Fast costume builder.
• Episode 20: AI-made scenario videos for compliance training.
• Episode 40: Sam has Claude Design write Veo 3 prompts for ad concepts.

FAQ

How much does Veo 3 cost per video?

When we built our video studio in June 2025, Jordan put Veo 3 at about $6 per eight-second generation. By October 2025 he remembered it as $6 to $8 a video and trending down, and he called Veo 3.1 Fast pretty inexpensive.

How long does Veo 3 take to generate a video?

In Episode 1, Jordan said Veo 3 sometimes took up to ten minutes for an eight-second clip, which is why he skipped running it live. ByteDance's Seedance took just a few moments for a similar prompt.

Is Veo better than Sora?

In our tests, yes. Sam thought Sora 2 trailed Veo 3 on physics in active scenes, and when we ran the same capybara prompt through both, the Veo 3.1 Fast clip looked better. Sora 2's edge is putting your own face in the video.

How do you use Veo 3 without a Google subscription?

We call it through Replicate's API from small apps built in Replit, so the whole team pays per generation instead of per seat. The same route powers our Halloween costume video app.

Can Veo generate dialogue?

It tries. In our Halloween app, characters say something like Happy Halloween, but Jordan noticed the rest often comes out as mumbled non-words.