OpenAI launched its ChatGPT Superapp this week, merging ChatGPT, Codex, and browser control into a single application. The app can take over your computer and browser to complete tasks end-to-end: editing video, writing production-grade code, building websites, generating spreadsheets. Andrew Ambrosino, OpenAI's lead for Codex and the ChatGPT app, told Big Technology: 'Anything that you can do on a computer, you can now do with this app.'
The app ships with a toggle between two modes: Codex, for users who want to dig into code directly, and ChatGPT Work, a less technical interface comparable to Anthropic's Claude for Work. Ambrosino is explicit that the toggle changes the experience, not the capability. OpenAI is also deliberately engineering against over-refusal, the pattern where AI agents decline actions like booking with a credit card. The system uses permission requests rather than hard blocks, and years of accumulated ChatGPT chat logs inform how the agent acts on your behalf.
The piece is worth reading in full for Ambrosino's explanation of why OpenAI chose a manual toggle over intent-detection ('there are a lot of use cases where we can't necessarily tell what you want'), and for the specifics on how memory from prior ChatGPT sessions feeds into the agent's real-world actions. The memory architecture is where the long-term competitive advantage lives, and the original reporting gets into the mechanics.
[READ ORIGINAL →]