BreakingTech retrospective archive — August 20, 2026 update.
On August 20, OpenAI updated ChatGPT and Codex with a series of features that clearly signal the product’s direction: less of an isolated window, more computer-level integration. On Macs with Apple chips, the Messages plugin can read and search iMessage, SMS, and RCS conversations, and can draft or send messages via the system app.
By default, sending requires user approval regarding content and recipients. This is an important detail because it illustrates the model we are likely to see more and more often: AI drafts the action, but certain sensitive operations still require explicit confirmation.
Computer History arrives in Europe as well
In the same update, OpenAI rolled out Computer History to the EEA, Switzerland, and the United Kingdom for Pro users on macOS. The feature allows ChatGPT and Codex to reference selected activities carried out on the computer, sparing users from having to rebuild context every time.
Computer History logs interaction events, not screenshots or audio, and is turned off by default. Users can choose which apps and websites to include and manage timeline items.
Memory becomes ambient
Until now, assistants have primarily remembered what happened inside a chat. The next step is remembering work that happens outside the chat: opened files, used applications, recent activity, and communications.
This is where AI can become truly useful in day-to-day work, but it is also where privacy risks increase. More context means better responses and fewer repetitive instructions; however, it also means granting the system broader visibility into one’s digital life.
From assistant to operating layer
Integrating Messages is a small example of a larger shift. When a model can read, search, write, and act inside applications, it ceases to be just a product you visit and becomes a layer connecting other products.
Future competition could therefore hinge on the quality of integration with the operating system and apps. A smarter but isolated model may be less useful than a slightly less capable model that understands context and can actually get the job done.
