GPT-5.6 Has Three Names. Luna, Terra, Sol

OpenAI released three models named Luna, Terra, and Sol, which correspond to small, medium, and large variants of GPT-5.6. The naming reflects an inference cost architecture based on the roofline model from high-performance computing, optimizing for compute-bound or memory-bound scenarios.

Member-only story GPT-5.6 Has Three Names. Luna, Terra, Sol OpenAI just released three models and named them after celestial bodies. That is not branding. That is an inference cost architecture announcement dressed up as a product launch. Let me show you the why. I’ve spent the last 48 hours pulling apart what these three names actually mean for anyone building production AI systems. Luna is the small, fast model. Terra is the medium. Sol is the large, capable one. Pretty names. But behind them is a cost optimisation problem that any systems architect should recognise immediately — because it maps directly to the roofline model from high-performance computing, a framework that has been around since 2009. The branding is the distraction. The math is the announcement. Grab coffee. This is going to take a while. — - Chapter 1: The Roofline Model, Translated for LLMs The roofline model asks one question: is your computation limited by how fast your hardware can do arithmetic, or by how fast it can move data? Define arithmetic intensity I as: I = FLOPs executed / bytes transferred If I is large — you’re doing a lot of computation per byte moved — you’re compute-bound. Faster math units help you.