Skepticism regarding Mythos inference optimization claims

scaling01 · x · 2026-08-15

The author doubts the rumor that OpenAI optimized inference for the Mythos preview, making it 2.5x cheaper to serve. The argument is that this implies the model was trained inefficiently, wasting billions of dollars, and then served internally for two months in that unoptimized state, which seems unlikely.

Original post →

More from Infra

Infra channel →