The interface is not freeform HTML. OpenAI describes a library of native, streamable components and a compiler that draws the screen as the model generates it, on web and mobile. The lab also says there is still work to do on the model’s design judgment. Both Sol and Luna are tuned for everyday conversation. The update applies to Chat.
On speed, OpenAI says GPT-6 interleaves thinking and answering. For web-search questions, GPT-6 Instant starts the answer 44 percent sooner on average than GPT-5.6 Instant. The lab says GPT-6 Extra High can start answering in the same time as GPT-5.6 Medium while scoring above GPT-5.6 Extra High on an internal set of everyday agentic tasks. An internal evaluation of difficult problems found GPT-6 addressing the key part of the question more often than GPT-5.6. Those comparisons are internal.
On safety, OpenAI says GPT-6 builds on Astra’s work, is clearer about the limits of its own knowledge, handles high-risk misuse more carefully than GPT-5.6 Sol, and resists multi-turn attempts to bypass safety training more effectively. The lab points to a system card for the measurements.