OpenAI Codex Review

Byline: The AI Take editor
Review scope: My hands-on impression, checked against OpenAI's current product documentation on 24 September 2026. This is not a controlled benchmark, a scored rating, or a comparison of equivalent tasks.

I have had good experiences using Codex Computer Use. The phone side has been less smooth: sometimes a message I send from my phone does not appear or sync where I expect it. That is my user-experience report, not a measured failure rate or a claim that every Codex user will see the same issue.

Codex's ability to work with a desktop interface is a real positive for me. When I steer a task remotely, though, the phone is part of the same agent conversation. A follow-up that does not show up when I expect interrupts that flow.

What OpenAI documents

OpenAI says Computer Use in the ChatGPT desktop app can let Codex see and operate graphical interfaces on supported macOS and Windows systems. The docs describe it as useful when a task depends on a desktop app or visual state that files and command output cannot show. On Windows it runs in the foreground, so the task takes over the active desktop while it works. These are documented capabilities, separate from my favorable impression of using the feature.

OpenAI's Remote connections guide says the ChatGPT mobile app can start or continue Codex chats on a connected Mac or Windows computer, send follow-up instructions, approve actions, and review results. The connected host supplies the project and tools; it must stay awake, online, and signed in to the same account and workspace. OpenAI also documents troubleshooting for missing hosts and approval requests. My occasional missing or out-of-sync message experience does not establish the cause or show how widespread it is.

How it feels to me

I like Codex's computer-use capability, but I have not found its phone conversation as dependable in practice as I would want. I have not measured how often the sync issue happens or identified a cause. Treat it as a reason to pay attention in your own setup, not as a confirmed product-wide defect.

For work I do mostly at the computer, Computer Use is a reason to keep Codex in consideration. If your workflow depends on steering an agent from a phone, check that messages appear reliably in your own setup.

Verdict

Codex has useful computer-use capability in my experience, while its phone handoff has occasionally disrupted the conversation. I report those observations without turning them into a universal product ranking. I still prefer Claude Code's interface overall because conversational ease and phone use carry more weight in my day-to-day work.

Sources checked 24 September 2026: OpenAI's Computer Use guide and Remote connections guide. Availability and setup can vary by operating system, rollout, account, and workspace. The hands-on impressions above are mine; they are not claims made by OpenAI.