We tested ChatGPT Work on two demanding jobs that go beyond normal chatbot demos: making and publishing a YouTube video from a minimal prompt, and finishing a half-built mobile app through the Apple Developer publishing flow.
The first test used ChatGPT Work running GPT-5.6 Sol at ultra-high effort. We gave the agent access to the local system, SeaDance 2.0, FishSpeech, CapCut Desktop, a Chrome plugin, and YouTube, then asked it to create a children's video with a minimal prompt and publish the final result. After roughly three hours, three human interventions, and more than three million tokens consumed, the video was created and published to YouTube: watch the result.
That result is impressive because it shows ChatGPT Work can keep going across multiple surfaces instead of stopping at advice. It handled creative generation, voice generation, editing workflow, browser work, upload steps, and publishing. Older AI assistants can usually write a script or suggest upload metadata; this test showed a work agent actually pushing through the production chain.
The second test was more practical. We asked ChatGPT Work to finish a half-built mobile app using an implementation plan and roadmap, with access to Apple Developer and EAS. The agent completed the app work and moved through the Apple Developer process toward App Store publication. At the time of review, the app is still in Apple review, so we will update this listing once the App Store link is approved.
The biggest strength is persistence. ChatGPT Work feels less like a chat window and more like a supervised production assistant. It can operate across local apps, browser sessions, developer tools, and publishing surfaces when the right permissions are available. For creators, founders, and small teams, that can turn stalled work into finished output.
The biggest weakness is operating cost and supervision. The YouTube test was not cheap or instant: it took around three hours, needed three human interventions, and consumed a very large token budget. We would not treat it as a fully autonomous worker for public actions. It is better understood as a capable agent that should pause for human approval at important checkpoints.
This matters especially because ChatGPT Work becomes more powerful when it has local-system access and account access. That is useful, but it should be scoped carefully. YouTube publishing, App Store submission, customer messages, production deploys, and paid campaigns are all workflows where the right behavior is not "never ask." The right behavior is "ask before irreversible or public actions."
Strengths: Strong cross-app persistence, useful for long-running work, can operate across local tools and browser workflows, strong fit for video, voice, and app-publishing tasks, good at turning clear plans into finished milestones, much more capable than a normal chatbot for messy multi-step jobs.
Weaknesses: Expensive for heavy tasks, can consume millions of tokens, still needs human steering, workflow quality depends heavily on tool access and account state, public publishing still requires review, not suitable for users expecting cheap one-shot automation from vague instructions.
Final verdict: ChatGPT Work is one of the first AI work agents that feels capable of completing real cross-app jobs instead of merely explaining them. It is not a replacement for skilled operators yet, but it is already a serious force multiplier for teams with clear goals, connected tools, and human approval checkpoints.