Every model’s fighting weight — who made it, how smart it is, what the API costs, and what it takes to run at home.

Score is the Artificial Analysis intelligence index; LIE (Local Intelligence Efficiency) is that score divided by the GB of RAM to run the model locally — intelligence density, higher is better; API prices are live from OpenRouter, USD per 1M tokens; Launch is the OpenRouter listing date. RAM is theoretical at 4-bit quantization (0.6 GB per billion parameters) — “unknown” means the lab hasn’t disclosed a size. Mac math: Mac Mini = 64 GB M4 Pro, Mac Studio = 256 GB (today’s top config), 75% of RAM usable; Power is the Mac Studio setup running 24/7 (200 W each) at $0.17/kWh. The RAM filter shows models that fit inside 75% of the chosen memory.

A Doranalytics project.