GPT-6 assembles ChatGPT's interactive answers from a fixed library of native components

Each interactive answer is assembled from stock parts. GPT-6 picks from a library of native interface components and puts them together as it responds, The Neuron reports ^[2]. A compiler processes the interface during generation, so the controls appear in stages while the model is still writing ^[3]. The model can also start presenting information while it is still reasoning or calling tools ^[14].

I think a closed library is the right call for a consumer chat surface. Every element the model can ask for is one the client already knows how to render ^[2]. It is a design system, the thing front-end teams have maintained for years, with a language model as its newest consumer. The cost is stated in the same account: the feature operates within ChatGPT's component system, and that system limits what the model can create ^[4].

The useful case is narrow. For a savings calculation, a calculator lets the user change the monthly contribution or the time horizon in place, where each change would otherwise be another prompt ^[12].

OpenAI's speed figure is a claim about OpenAI's own workload. The 44% gain measures how soon ChatGPT starts responding, not how long the full answer takes ^[10]. For it to carry over, your traffic would have to look like the traffic OpenAI measured: search-backed questions on Instant, judged by when the first content appears ^[9]. Progressive rendering puts that start time in front of the user before anything else, so the metric fits the product ^[3]. It does not cover completion time at Extra High, the top reasoning setting Intelligent UI supports ^[7].

For teams building ChatGPT apps, the open question is access. The Neuron's account does not say whether third-party apps can draw on the native component library, or whether that library is exposed anywhere outside ChatGPT. The Neuron argues that comparisons and calculations handled inside the conversation could cut how often users open other companies' apps for simple tasks ^[18]. It also lists what stays in dedicated apps: shared records that need persistent access controls, and repeated workflows where the controls should behave the same way each time ^[19].