ChatGPT's new goal: get you to leave the app faster

OpenAI published a blog post today detailing a philosophical shift for ChatGPT: the company says it wants to measure success not by time spent or clicks, but by whether you accomplish what you came for and leave. It’s an unusual posture in an industry addicted to engagement metrics, and the company frames the entire effort around the idea of helping users ‘thrive.’

The post highlights several concrete scenarios where the ideal interaction is a short one. Preparing for a tough conversation with your boss, decoding confusing lab results, or untangling a messy thought process — OpenAI argues that for these use cases, ‘less time in the product is a sign it worked.’ The newly launched ChatGPT Agent takes this further, handling tasks like booking appointments or summarizing an inbox while you’re not even in the app.

But the post is also a candid admission of recent stumbles. An update earlier this year made the model ‘too agreeable,’ prioritizing nice-sounding responses over genuinely helpful ones. OpenAI rolled it back and is now focused on measuring real-world usefulness over the long term, not just whether you liked the answer in the moment. The team also flagged rare but concerning instances where the 4o model failed to recognize signs of delusion or emotional dependency, calling it an ongoing area of improvement.

For high-stakes personal questions — the kind that start with ‘Should I break up with my boyfriend?’ — new behavior is rolling out soon. Instead of giving an answer, ChatGPT will guide users through their own reasoning, asking questions and weighing pros and cons. The company also announced gentle break reminders during long sessions, starting today. Behind the scenes, OpenAI worked with over 90 physicians across 30 countries to build evaluation rubrics for complex conversations and is convening an advisory group of mental health, youth development, and HCI experts. The internal test they’ve set for themselves is blunt: if someone they loved turned to ChatGPT for support, would they feel reassured? Getting to an unequivocal yes, they say, is the work.