GPT-6 Astra One-Shots Private Humor Eval That Every Other Model Has Failed

mattbeane · x · 2026-09-05

Researcher mattbeane reports that GPT-6 Astra one-shotted his private eval — a test every model had failed until now.

The entire prompt was just "why is this card funny?", requiring the model to grasp the humor behind an image. It hints at a generational jump in humor and social common-sense reasoning, though it's a single sample and needs broader validation.

Original post →

More from Models

Models channel →