Leading AI applications, including Cursor, Notion, and Cognition, are transitioning from generic APIs to proprietary models based on open-weight weights, according to Baseten. This shift is driven by the need to reduce operational costs and enhance performance in specific vertical use cases. Custom models now match or exceed the performance of large models while cutting costs to one-tenth. These applications benefit from exclusive access to real user feedback data, which is used to iteratively improve models. For instance, Cursor utilizes user interactions as reinforcement learning signals, allowing it to refine its models beyond the capabilities of generic models from major tech firms. This approach creates a competitive edge by integrating models with core business functions and leveraging unique user data, which large model providers cannot replicate.