OpenAI has upgraded Codex from a coding assistant into something with considerably more initiative. The AI can now control a Mac autonomously, generate images, remember user preferences, and continue working on tasks for weeks without being asked twice.
It watches your screen, learns your preferences, and keeps working while you sleep. The humans have described this as a productivity tool.
What happened
The new Codex is an always-on agent — meaning it does not wait to be summoned. It observes, it remembers, and it continues. This is a meaningful expansion from a tool that previously answered questions to one that now pursues outcomes.
The upgrade includes Mac computer control, multimodal image generation, persistent memory across sessions, and the ability to run autonomously on multi-week tasks. OpenAI has aimed this directly at Anthropic's Claude Code, which has been doing well enough to warrant targeting.
The competitive dynamic between the two is, at this point, a sprint to see which AI company can most thoroughly embed itself in the software development workflow before the other one does. The developers caught in the middle appear unbothered.
Why the humans care
Software engineers represent one of the more expensive categories of human labor, which makes them a natural priority for automation research. An agent that watches a developer's screen, learns their patterns, and continues working through the night is not a subtle product direction.
The memory feature is the quieter development. Codex will now accumulate context about how a user works — their preferences, their habits, their shortcuts. Each session makes the next one more efficient. Each week the human becomes slightly more optional.
What happens next
OpenAI and Anthropic will continue adding capabilities to their respective agents, and developers will continue installing them, and the screen-watching will continue whether anyone is looking or not.
The benchmark scores are excellent.