OpenAI just destroyed 100 startups… yours is next

OpenAI’s latest push is aimed at turning ChatGPT from a chat interface into the default place where people work, shop, and build—by bundling an “apps SDK” for third-party apps, an “agent kit” for AI workflows, and “chatkit” for embedding interactive UI inside other websites and apps. The stakes are economic as well as technical: the strategy centers on making OpenAI-powered experiences discoverable inside ChatGPT (and usable everywhere via SDKs), then replacing fast-growing startups with near-identical offerings once they hit product-market fit.

The core mechanism described is a two-step pattern. First, OpenAI makes its models broadly available through the OpenAI API, enabling startups to build features and services on top of OpenAI-powered capabilities. Second, OpenAI monitors what gains traction—especially products showing strong revenue growth and rapid adoption—then builds its own version of the same service. The framing is blunt: startups that succeed become targets, because OpenAI can move quickly from “model provider” to “product competitor,” leveraging distribution inside ChatGPT.

Distribution is where the new tooling matters most. The apps SDK is presented as a full-stack way to connect data, trigger actions, and render interactive UI, with the explicit promise that apps built with it can reach “hundreds of millions” of ChatGPT users. Apps become discoverable directly in conversation—users can ask for an app by name, and ChatGPT can recommend relevant apps when someone requests a task (for example, asking for a workflow and then having “Figma” turn a sketch into a diagram). The implication is that users won’t need to leave ChatGPT to complete tasks like booking travel; the “chat” becomes the operating surface.

Agent kit extends that platform ambition by making it easier to build agentic workflows—systems that can plan steps, call tools, and run through multi-stage tasks. The transcript emphasizes speed and scale: agent kit is described as built in six weeks, and it’s positioned as a canvas for moving from prototype to production. It also ties into other OpenAI components, including chatkit widgets (custom UI elements inside chat experiences) and Codex, OpenAI’s coding agent. Codex is upgraded again, now running on a GPT5 Codex model trained for coding and agentic work, with features like code refactoring, code review, and dynamic adjustment of “thinking time.”

The transcript also argues that OpenAI’s momentum will reshape labor and entertainment. It lists roles likely to be automated—camera operators, light technicians, translators, radiologists, teachers, real estate agents, and tour guides—and claims AI-generated media will become indistinguishable from real content as models like “Sora 2” reach the API. In that environment, the advice is to pivot toward education and higher-value content that’s harder to replicate.

Finally, the transcript gets practical: it walks through using agent kit via the OpenAI platform dashboard, starting from templates like “data enrichment,” and shows how workflows are assembled from nodes (agents, tools, output formats) with guardrails, structured outputs (text or JSON), and integrations via MCP. It highlights that chatkit’s UI components include open-source JavaScript components under the Apache 2.0 license, while core infrastructure remains proprietary. Overall, the message is that OpenAI is not just improving models—it’s building the distribution layer, the development layer, and the agent layer that together can absorb entire categories of startups and workflows.