Choosing Work over Chat in paid ChatGPT spends more quota on OpenAI's top models

A baker wants ideas for new kinds of Kouign-amann and, in the same sitting, a purchase order for flour. In paid ChatGPT those now belong to different modes, Chat for the ideas and Work for the PO, according to a Fast Company writer who describes being a heavy, perhaps addicted, user ^[13][17]. The writer warns that the wrong setting can derail a project, or at least cost you some of ChatGPT's most powerful functions ^[2].

In Chat, that user asks whether phones are allowed during jury duty, and whether a raccoon pooped in the garage ^[16]. The writer makes the call once, on sitting down, by asking whether the session should produce information or something tangible ^[14].

People notice the personality change first. The writer found Work "far more gruff and business-like" and Chat more collaborative and friendly ^[8][3]. The part a budget owner cares about sits underneath. Work typically runs OpenAI's most advanced models and thinks longer before answering, and the writer warns it can use up a usage quota very quickly ^[10][12]. On the other side, the writer wrote: "No one wants to wait seven minutes for a response when asking a basic question." ^[6]

The link tagging is the clearest signal of what each mode is for. In Work, the UTM tags that Chat attaches to outgoing links are generally missing ^[10]. The writer takes that to mean the bot does not expect you to follow its work with research of your own. "In short, Work mode turns ChatGPT into what's basically an agentic AI platform," the writer wrote ^[11].

These are one person's field notes. The writer hedges with "often", "typically" and "generally", and the article does not include quota figures or a statement from OpenAI ^[18].

The forcing question for a rollout is who takes the next step once the session ends. If a person will click through and decide, it is a Chat task. If the output is the handoff, such as a table or a PDF report, it is a Work task ^[14]. Add a second axis, how often the task repeats, and you get four cells. A frequent question with a human next step goes to Chat by default. A one-off deliverable goes to Work. A recurring report, like the monthly content audit in the writer's test, is a file that repeats, so it gets Work plus someone watching the quota ^[9][10]. One-off vendor research ends in a person's decision, so it stays in Chat even when it feels important enough for the bigger model. The writer also picks Chat when a more personable tone is wanted ^[15].