Jordan builds a Sora 2 API video tool, tests it against Veo 3.1, then shows an n8n sales outreach workflow and a Claude-generated n8n workflow in minutes.
In a quick side by side test, Google's Veo 3.1 fast made more realistic video than OpenAI's Sora 2 from the same prompt, and both are now available to build your own tools on. On Episode 17 of Built This Week, Jordan Metzner and Sam Nadler showed a video generator Jordan built on the new Sora 2 API, compared it with Veo 3.1, and then walked through how we build n8n automation workflows, including one Claude wrote from a single prompt.
Jordan says the name Sora causes some confusion. Sora 2 is OpenAI's new video generation social app, and it is also the name of the video model you can call through OpenAI's API. The API is the more interesting part to him, because you can build your own app around it and make any kind of content. He sees reference images, where you upload an image and the model builds the video from it, as one of the first ways to get the continuity you need to start making movies.
He built a small Sora 2 tool for our teams. It generates horizontal or vertical videos, lets you choose the length, accepts a reference image and can remove the Sora watermark, a feature Sam called a great little touch for anyone posting to their own channels. Jordan also shared the rough edges: the reference image has to be an exact size or it fails, and very long prompts sometimes make the API fail even with stop gaps in place.
Watch the Sora 2 vs Veo 3.1 test in Episode 17 (starts at 4:54) · Episode 17 notes and transcript
Google released Veo 3.1 the day before we recorded, and Jordan had already added Veo 3.1 fast to our internal video generator, calling it pretty good and pretty inexpensive. He ran the same prompt through both models: two capybaras talking on the Built This Week podcast, saying like and subscribe, with Sam's addition of something unexpected and funny at the end, at 12 seconds.
Jordan's verdict was that the Veo 3.1 version was better. He said the capybaras looked super real, the kind of shot that would take a ton of CGI, while the Sora ones did not look as real but were still pretty good. Sam agreed that Veo 3.1 looked better, and pointed to the release pace: two weeks earlier everyone was talking about Sora, a few months before that it was Veo 3. We keep a running comparison on our Sora vs Veo page, and we first tried Sora 2 in Episode 15.
Each clip took only a few minutes. Jordan's point was that you can send the same prompt to Sora and Veo several times, get different results, and stitch the best ones into a short film, trailer or TV commercial. We had already started doing that for our own businesses.
So I think, like, this is really the beginning of content generation, automation of video at scale.
— Jordan Metzner, Episode 17
He called it a huge threat to Hollywood and to creators who film content the traditional way, but also said it democratizes content creation. A kid in his bedroom could never have made a capybara TV commercial before. Now he can prompt one.
Jordan describes n8n as a platform for agentic workflows, or workflows in general, and says it is probably not that different from Zapier and Make. What n8n did, in his view, was build the agentic workflow pattern that other players, including OpenAI, are now copying. It connects to most major tools, such as Gmail, Google Sheets and Airtable, and can call third party APIs like OpenAI and Perplexity.
Sam says it is meant to be low code or no code, but there is a learning curve and it can feel daunting at first. Our shortcut:
But we've found a way to use Claude to build out, 90% of really any automation you could think of and then just kinda tie it all together at the end.
— Sam Nadler, Episode 17
Jordan showed one of the workflows we built, which he runs on a daily schedule and triggered by hand for the episode. It works like this:
The run includes wait steps so it does not spam any APIs, and n8n shows the logs and how many tokens each step uses as it calls OpenAI. The full workflow is written up on our n8n sales outreach build page.
Yes, at least the structure. Sam's method is to ask Claude for an n8n workflow, take the JSON it produces and import it, then spend the remaining 10% connecting the accounts, which he admits can be challenging. Jordan tried it live with a loose prompt for a workflow that emails podcast hosts to invite them on Built This Week.
He downloaded the JSON, imported it into a new n8n workflow from a file, and it came in with a schedule and manual trigger, a step to get podcast leads, filtering, data prep, a fetch of each podcast's website, email generation and parsing. Sam said it is not a final version, but it is pretty good, and it is how we built several workflows in a couple of weeks. Jordan also tried the ChatGPT agent workflow tool and called it very rudimentary by comparison.
Sam brought up Walmart partnering with OpenAI so shoppers can buy directly in ChatGPT, with Amazon not part of it. Jordan thinks most major apps will integrate because it is basically free traffic, and that chat is becoming the new search interface. People already start searches in ChatGPT; when his pool pump broke, he could have photographed it and bought a replacement without leaving the chat.
His concern for Amazon: Google has long driven traffic to Amazon, and Amazon spent twenty years getting shoppers to start product searches on its own site. If those searches move to ChatGPT, Amazon could lose conversion and may have to pay ChatGPT for traffic the way it has paid Google. He compared the moment to the early 2000s web, when everyone was doing deals with everyone.
Jordan's take was lukewarm. He says that, unlike earlier generations, Apple did not announce premium versions of the new chip, so if you want the top MacBook Pro with the most RAM, you still have to buy an M4 Pro. He also said there is little reason to buy a base M4 now that a base M5 costs the same.
The pitch is running LLMs locally, but Jordan says we do not really do that, since we mostly use managed model services, and when he moved to M4 with a lot more RAM he did not notice a difference. He called it a so so launch rather than a superstar one.
And people care about features. They don't care about performance.
— Jordan Metzner, Episode 17
The episode also covers Anthropic's revenue growth and how much our own work depends on OpenAI and Anthropic. Watch the capybara clips and the live n8n runs in Episode 17 of Built This Week, where the full transcript is also on the page.
