June 22, 2026. Bubbles, computational complexity, and my experience with Fable 5.

Introduction

Like many people around the world, I was excited to get my hands on Anthropic’s frontier model, Fable 5, when it came out. For a little while now, my test of frontier models has been simple: to formulate and prove a rigorous version of an intuition I have about soap bubbles. Opus 4.8, GPT 5.5 and Gemini 4 (Pro) all fell short, either proving something erroneous, or mismatching result and intuition. By bouncing Fable 5 off another model, however, it gave something surprising, novel, and seemingly correct.

Bubbles

Long ago, I wrote a semi-popular monograph on the links between soap bubbles and computational complexity.