Creators praise Codex in ChatGPT mobile app, say it outperforms other coding agents

OpenAI added Codex to the ChatGPT mobile app in preview on May 14, giving users a way to monitor, steer and approve coding tasks from a phone while work continues on a laptop, devbox or remote environment. In the same week, YouTube creators published a cluster of comparison videos that cast Codex less as a benchmark entrant than as a day-to-day coding agent measured on repo work, task routing and long-running execution. OpenAI said in its product post that more than 4 million people now use Codex every week. The company described the mobile app as a way to review outputs, approve commands, change models and start new work across active threads. ### Why are creators suddenly talking about Codex on phones? OpenAI said on May 14 that Codex is “now in preview in the ChatGPT mobile app,” extending a product it had previously pitched as a desktop command center for multiple coding agents in parallel. The company said the mobile version loads the live state from the connected machine, with screenshots, terminal output, diffs, test results and approvals flowing back to the phone in real time. (openai.com) The shift matters because OpenAI has been presenting Codex as an agent workflow product rather than a single chat box. Its Codex pages describe built-in worktrees, Git support, cloud environments and parallel threads, while release notes say remote access now lets users answer questions, redirect execution and switch between connected hosts from ChatGPT mobile. ### What exactly are the YouTube creators claiming? (openai.com) A YouTube video posted about 20 hours before this report was titled “This FREE AI Coding Agent is Insane | Better than Claude Code,” explicitly comparing a free alternative with Claude Code and Codex. Another video published about 12 hours earlier in German was titled “ChatGPT war gestern: Codex ist die beste KI-App von OpenAI (nicht nur fürs Programmieren),” framing Codex as OpenAI’s best app and not only a programming tool. (openai.com) Those titles do not amount to independent performance verification, but they show how creators are positioning the category. The comparison point in both cases is not a synthetic coding benchmark cited in the title or description. It is whether the agent can handle larger tasks, reduce prompt back-and-forth and stay useful across an existing repository or broader workflow. That reading is an inference from the videos’ framing and accompanying descriptions. (youtube.com) ### Where does Paperclip fit into this? Paperclip is being pitched by creators as a controller layer for multiple agents. A recent tutorial video described it as a foundation for “hiring” multiple agents such as OpenClaw, Claude, Codex and Cursor, assigning them roles and managing workflow through a command structure. Another video description said users could switch to Codex mid-session after running out of Claude tokens. (youtube.com) That overlaps with OpenAI’s own product language. OpenAI has described the Codex app as a command center for agents and said multiple agents can work in parallel on a Mac without interfering with other apps. The similarity is not that the products are identical, but that both are being described around orchestration, routing and supervision rather than one-shot prompting. ### Are creators moving the debate away from benchmark scores? (youtube.com) OpenAI’s recent Codex materials emphasize implementation, refactors, debugging, testing and long-running tasks, and the company’s changelog highlights model switching inside the app. Creator videos surfacing this week mirror that emphasis by talking about practical repo work, free usage, controller apps and comparisons with Claude Code in ordinary development loops. (openai.com) The result is a different kind of comparison set. Instead of asking only which model is strongest on coding benchmarks, creators are asking which agent is easier to supervise, which one can keep several threads moving, and which setup lets developers route work between tools. That conclusion is based on the titles, descriptions and official product pages available as of May 19. ### What should readers watch next? (developers.openai.com) May 19 is still within the first week of Codex mobile preview, and OpenAI’s help and release pages say setup requires updated ChatGPT mobile and Codex apps, with the host machine awake, online and running Codex. The next concrete signals are likely to come from additional creator tests on repo tasks, multi-agent routing and mobile approvals as the preview expands. (help.openai.com) (youtube.com)