Higgsfield, the AI video ad platform, published what they're doing with GPT-6 Astra, and one line stopped me. A customer can ask for a hundred variations of their top performing ad and get them. Different creative directions, including versions localized for different countries. 🎬 Their CEO also said a single engineer is now shipping new features in a day, which he puts down to the model planning work across many steps instead of handling one task at a time. So here's where we are. Last week I made a 38 second product video with Opus 5.5 writing the actual frames in code. This week a platform is generating a hundred ad variants from one prompt. Production cost for creative is heading toward zero, fast. Which sounds like great news, and mostly is, but it leads somewhere uncomfortable. 💡 Making 100 ads was never the hard part. You could always have hired three freelancers and had a hundred variations in a week if you really wanted them. Volume was expensive, but it was available. The hard part has always been knowing which three to actually put money behind. And that doesn't come from a model. It comes from knowing the customer, reading the performance data honestly, and having been wrong enough times that you recognise the pattern. So the constraint moved. It used to be production capacity. Now it's judgment and testing discipline. What that means practically. If you sell creative, stop pricing per asset. The asset is about to be worth very little and your client will figure that out. Sell the selection, the testing structure, and the call on what to scale. If you buy creative, the same thing in reverse. Paying per video is about to look silly. Pay for whoever can tell you which video deserves budget. And if you run ads for your own business, the real upgrade here isn't making more. It's that you can now afford to test properly. Running twelve genuine variants instead of two was previously a budget decision. Now it's a Tuesday. One caveat worth stating. That writeup is published by OpenAI about a company using OpenAI, so treat it as a case study rather than a neutral test.