OpenAI updates Codex: the agent can now use your computer, remember context and plan ahead

OpenAI has substantially expanded Codex, effectively moving it from the "coding assistant" category into the class of full-fledged working agents. After the update of April 16, 2026, the service can not only write and review code, but also work with applications on a computer, remember user context, use a built-in browser and continue long tasks without constant human involvement.

For OpenAI, this is an important move in its direct competition with Anthropic, whose Claude Code many consider a more mature tool for agentic development. The company says Codex is already used by more than 3 million developers per week, and OpenAI is now trying to establish it not only as a tool for programmers but also as a universal assistant for product, research and operations teams. The logic is clear: if an agent can understand code, documents, comments, spreadsheets and external services, it becomes useful far beyond the IDE.

The main change is background work with desktop applications. Codex can now see the interface, click with the cursor, type and open programs, performing actions the way the user would. At launch, OpenAI is rolling this mode out on macOS and stresses that several agents can work in parallel without interfering with what the person is doing in other windows. The practical point is simple: you can hand the agent routine tasks such as checking a frontend, re-running scenarios, navigating internal tools or acting in programs that have no API at all.

The app also now has its own browser. In it, the user can leave instructions directly on a page, and Codex can carry out actions in a specific web interface. This is especially useful for frontend development, interface testing and fast design iteration. OpenAI has also added image generation, so a single workflow can produce not only code but also mockups, diagrams, slides and visual sketches. Another notable block is integrations: Codex now has more than a hundred connections to third-party tools, including services for code review, task management, CI/CD and enterprise work.

The update also touches the developer's usual working setup. Codex can now process comments in GitHub review, work with several terminal tabs, connect to remote devboxes over SSH and open not only code but also PDFs, spreadsheets, presentations and documents in the side panel. A separate summary panel shows the agent's plan, the sources it used and the artifacts it collected. This detail matters: OpenAI is trying to turn Codex from a chat for one-off requests into a single workspace where the agent writes code, checks the result, prepares edits and shows the intermediate state of its work.

Another layer of updates concerns long-running tasks. Codex gets a memory mode in which it remembers the user's preferences, past fixes and context it has already found. The agent can also reuse old threads, plan the continuation of its work itself and "wake up" later to return to a task hours, days or even weeks on. In practice, this opens up scenarios such as a daily digest from Slack, Notion, Gmail or Google Docs, as well as automatic follow-up on pull requests and work discussions.

In parallel, on April 2, 2026, OpenAI moved Codex billing to a more flexible pay-per-actual-usage model, which makes the tool more convenient for teams with irregular workloads.

The point of the update is that Codex looks less and less like an ordinary code generator and more and more like an operating shell for digital tasks. If OpenAI can maintain reliability, transparency of the agent's actions and a reasonable cost, Codex may become stronger precisely where the company previously lost to Anthropic: everyday practical usefulness. But the competition is no longer only for programmers; it is for the place of the main interface for team work with AI.