The claim that AGI has arrived usually rests on a lowered bar: a system that can discuss anything must be general. That conflates breadth of training data with breadth of capability. Ask such a system to acquire a new skill after deployment, or to complete a multi-day task without supervision, and the gap appears quickly.
What would count as arrival
A system that, faced with an unfamiliar domain, identifies what it needs to learn, learns it, applies it, and can explain its reasoning, reliably enough that a professional would accept the work. No publicly demonstrated system meets that bar today.
Why the question keeps returning
Because capability is improving quickly in visible ways, and because the definition of AGI has no enforcement mechanism. Each new model resets what casual observers think general means.
The useful reframe
Instead of asking whether AGI is here, ask which specific capability is missing and how you would test for it. That question has answers.