Hi,

OpenAI shipped a batch of Codex updates over May 21-22. There is no new model here, so I almost skipped it - but reading the release notes, a few changes stood out that matter for how the agent feels day to day.

One thing up front: this is a desk review based on the official release notes. I have not run Goal mode or Appshots hands-on. Treat the assessment as a read of the spec, not a test result.

This is not a new model - it is an update that firms up Codex’s foundations as an agent. I would call it a significant strengthening of existing features. The headline is Goal mode reaching general availability: hand over an outcome and success criteria, and let it run for a long time. Appshots and the Chrome-related improvements are not flashy, but they cut the setup overhead you pay on every task.

The catch: the underlying model has not changed. This is not an update to jump on expecting better output quality.

The May 21 ChatGPT release notes collect the Codex updates: Appshots, Goal mode going GA, in-app browser annotations, locked Computer Use, browser-use improvements, and plugin sharing.

Appshots - press a hotkey in the macOS Codex app to attach the window you are looking at, with a screenshot and extractable text, to a thread. No long preamble needed.

Codex CLI v0.133.0 (May 22) - Goals on by default with progress tracking, remote-control behaving like a foreground command, permission profiles, better plugin discovery, more lifecycle events for extensions. Mostly an ordinary minor release otherwise.

Goal mode going GA is the part I think matters most. What was experimental is now a normal feature across the app, IDE extension, and CLI. OpenAI says that once you define an objective, it can keep working for hours or even days (official announcement, retrieved 2026-05-23). Writing the outcome and success criteria up front should suit routine, well-defined work you want to hand off.

Appshots is modest but practical - if attaching what you are looking at becomes one hotkey, that is an action you repeat many times a day.

Locked Computer Use keeps work running after a Mac locks, with safety-oriented design: short-lived authorization, covered displays, re-lock on local input. It reads like a companion to long-running Goal mode.

Plugin sharing (ChatGPT Business) distributes locally built plugins within a workspace. Enterprise is off by default, Edu is on by default - check your plan if you are on a team.

Already using Codex? It is worth bumping the CLI to v0.133.0 and trying Goal mode once - longer refactors and fixed-step routine work are where I would expect a difference.

Not using Codex yet? This update alone is not a reason to switch. The model has not changed; if output quality is your deciding factor, wait for a model update. For now, read this as a maintenance release leaning toward long-running agent operation.

Again: this is written from the official release notes, retrieved May 23, 2026. No hands-on run of Goal mode or Appshots. Since it is not a new model, there are no new benchmark numbers. I will do hands-on verification on a future occasion.

ChatGPT Release Notes (OpenAI Help Center): https://help.openai.com/en/articles/6825453-chatgpt-release-notes

Codex changelog (OpenAI Developers): https://developers.openai.com/codex/changelog/

Codex CLI v0.133.0 (GitHub): https://github.com/openai/codex/releases