How Jordan built a shared AI video and image studio in Replit with Replicate, why Veo 3 costs add up fast, and how the team made a capybara ad in a day.
You can give a whole team access to the best AI video models without buying seats for each one by building a small internal studio on top of model APIs. That is what Jordan Metzner and Sam Nadler showed on the first episode of Built This Week: a shared video and image generation tool Jordan built in a few days with Replit and Replicate, which the social team was already using to make commercials.
Jordan says the idea came from a chat tool the company already runs for its people. Instead of paying for a subscription to each AI lab per person, staff pick from many models, from ChatGPT to Claude to DeepSeek, and the company pays through API keys as a consumption layer. You pay for what you use, and everyone gets access to more models.
Then Google launched Veo 3. Jordan says Sora had made some rumblings, but Veo 3 felt like a step function improvement. The social team, which produces video for several brands and for clients, immediately asked how they could get access. Jordan's answer was to do what they had done with chat: give them Veo 3 and as many other top models as possible in one place. Producing videos, from shoots to filming, had always been their biggest bottleneck.
Watch Jordan on what Veo 3 really costs in Episode 1 (starts at 6:19) · Episode 1 notes and transcript
Sam asked the obvious question: if Veo 3 is the best model, why not just give everyone Veo 3? Jordan's answer was price. At the time of recording, he said Veo 3 cost about $6 per eight-second generation, and the generations are not always usable. A 30-second commercial needs at least four clips if every output is perfect. In practice you might generate 10 clips and spend $60, $80 or $100 in a few minutes.
He was clear that this is relative. Compared with a full day of shooting, Veo 3 is much cheaper, and for a marketing agency making commercials for big brands, a few hundred or a thousand dollars on concepts is cheap. For a startup, though, the bill grows quickly:
But from a startup perspective, pretty easily you can spend hundreds if not thousands of dollars if you don't pay attention.
— Jordan Metzner, Episode 1
Jordan demoed the studio live. It lists six of the top video models, each with its own pricing and strengths, and he says the only way to know which one suits your use case is to generate samples and compare. You pick a model, set options like a reference image or frame rate, and type a prompt. They used "a dog skateboarding in Venice, Italy."
One of the first features they added was an enhance with AI button. It sends your prompt to ChatGPT along with the model you are using, and returns a richer prompt; here it turned the one-liner into a golden retriever skating past gondolas in warm afternoon light. Jordan says both prompts work, but the detailed one produces better output. Sam pointed out that a copy prompt button came next, so the team could test the same prompt across models.
Speed varies a lot. Jordan says Veo 3 can take up to ten minutes for an eight-second clip, which is part of why it costs more, while Seedance, ByteDance's newest model, finished in moments. Running the same prompt on Veo 2, Google's previous model, gave a scene that looked more like Venice but with lower video quality. He has since added an image page with four top image models, including Flux Pro and MiniMax, which are much faster than video. Everyone shares one account, so the team can see each other's history instead of each person needing seats at Runway or Kling.
Jordan built the studio in Replit. He says he reaches for Replit when a tool needs a back end, a database, storage for keys and a little authorization, but won't scale to a huge organization. For a front end only product demo he would use Bolt, Lovable or Cursor instead. His first prompt was probably something like a web admin panel with a left-hand sidebar that uses different models to generate video, sometimes polished in ChatGPT first.
The models come from Replicate, which hosts many models behind APIs (he also named fal and AWS Bedrock as alternatives). Each model page has a playground plus an API tab with a sample call and the schema. Jordan copies the sample call and schema into Replit's agent and asks it to add the model, and it usually works out of the box. Some models still failed during the demo, and he said he would keep fixing them. Every generation is a paid Replicate call.
Three or four days after release, the tool had already produced two campaigns. Jordan made a Wes Anderson style commercial that got good reviews on social. He thinks part of the reaction is that AI video is still new, so the more realistic it looks, the more surprising it is. Then one of their video editors made a National Geographic style ad for their offsite planning business, Offsiteio, about capybaras going on an offsite, with a voice-over in the style of Sir David Attenborough.
Our video editor, once given access to this tool, generated probably 20 to 30 clips, aligned with this concept. Had it done within, an hour maybe, stitched them together, layered on the voice over, presented it to us, few small edits, and within a day we had a commercial that previously would have cost, thousands of dollars or just been too time consuming or costly for us to even pursue.
— Sam Nadler, Episode 1
Jordan added that doing the same with CGI would mean modeling capybaras in 3D, placing them in environments and waiting on render farms. This took a few hours. We later built an AI video editor with Remotion, covered in Episode 36.
The tool of the week was Wispr Flow. Jordan types about 60 words a minute and found Apple's built-in dictation clunky. With Wispr Flow you assign a key, talk, and it types. He says the app measures him at about 160 words a minute, and it handles multiple languages, which matters to him because he speaks a lot of Spanish. He uses it for emails and for talking to coding agents like the one in Replit. At the time of recording he put the price at about $10 to $12 a month per user, and the company had just raised $30 million.
Sam has used it daily for months. His reason:
what I love about it is when you have to chat with ChatGPT or Claude and you have to provide a lot of context to your prompt in really any use case, the more context and guidance and structure you give the prompt, usually a better outcome. And typing that out can just require minutes and minutes of effort.
— Sam Nadler, Episode 1
He also likes the dictionary feature, which learns business names that a normal dictation tool would mangle. Asked whether Apple could make Wispr Flow redundant, Sam said Apple is fully capable of it, but whether it is a priority is another question.
In the news segment, Jordan pointed to a lawsuit that revealed Anthropic had bought used books, cut them up and scanned them. In his view it shows labs doing things that don't scale to get data that isn't already on the internet. Sam pictured a warehouse of people ripping pages from bindings and called it a race to capture data.
Jordan demos the video studio, the prompt enhancer and the Replicate workflow on screen in the full show. Watch Episode 1 of Built This Week, where the full transcript is also on the page.
