Sora 2 API Video Generator

Jordan built a small video tool on OpenAI's newly released Sora 2 API, less than a week after it came out, and showed it in Episode 17 (October 17, 2025). It generates horizontal or vertical clips, lets you set length, add a reference image and remove the Sora watermark, and keeps a video history. He called it "a quick Sora tool" but didn't give a build time. Then he ran the same prompt on Google's Veo 3.1 Fast, and Veo won.

Why we built it

Jordan wanted our teams to have the newest video model at their fingertips. He also pointed out at 02:18 that "Sora" means two things: OpenAI's video social app, and the model you can call through the API. The API is the interesting part, because you can build your own content tools on it, and reference images give you continuity between shots, "one of the first ways you can start to kinda make movies."

The stack

• Sora 2 API: the video model.
• Our existing Ryz Labs Video Studio: ran Google Veo 3.1 Fast for the comparison.

Jordan didn't say which coding tool he used for this one.

How we built it, step by step

1. Stand up a simple tool on the Sora 2 API. A separate app, since Sora 2 was brand new.
2. Add the controls. Orientation, clip length, reference image upload and a remove-watermark option.
3. Respect the reference image size.

When I add a reference image, it has to be this exact size. Otherwise, it doesn't work.

— Jordan Metzner, Episode 17

4. Guard against long prompts. Very long inputs sometimes make the API fail. Jordan added stop-gaps; it still happens occasionally.
5. Test it. The prompt: two capybaras on the Built This Week podcast saying "like and subscribe," with Sam's addition that something unexpected and funny happens at the end. Length: 12 seconds.
6. Run the same prompt on Veo 3.1 Fast in our main video tool, and compare.

How it turned out

Each clip took a few minutes. The Sora version was "okay." The Veo 3.1 Fast versions were funnier and more realistic: in one, the capybaras had bananas; in another, one fell asleep mid-show, which Sam said was "oddly accurate" for Jordan. Jordan's verdict: the Veo capybaras "look super real," and the Sora ones less so, but still pretty good.

Sam's takeaway was the pace. Two weeks earlier everyone was talking about Sora, before that Veo 3, and now Veo 3.1. Jordan's view: send the same prompt to several models, several times, and keep the best of each.

What we'd do differently

• Don't build a separate app per model. Sam asked why it wasn't just another option in our main video studio. Jordan agreed and planned to fold Sora in and retire the standalone tool.
• Validate inputs. Resize reference images to the exact size and cap prompt length before calling the API.

For a fuller comparison, see Sora vs Google Veo. The same episode covered our n8n sales outreach workflow.

FAQ

How do you build an AI video generator with the Sora 2 API?

Jordan built a standalone tool on the Sora 2 API with options for horizontal or vertical video, clip length (he used 12 seconds), a reference image and watermark removal, plus a video history. He added stop-gaps for API failures.

Can you remove the Sora watermark?

Jordan's Sora 2 API tool has a remove-watermark option built in. Sam liked having it in the tool for content going on our own channels, noting there are many separate tools that do this.

What are common Sora 2 API errors?

Two tripped Jordan up: reference images must be an exact size or the call fails, and overly long prompts can make the API fail. He added stop-gaps, but it still happened sometimes.

Is Sora 2 or Veo 3.1 better?

In our single side-by-side test with a capybara podcast prompt, both hosts thought Veo 3.1 Fast looked better and more realistic. Sam's point was that the lead changes every few weeks.