tl;dr: Open-weight AI models have essentially closed the capability gap with closed frontier models (3.3% on Chatbot Arena, at parity for coding), now route the majority of production tokens on OpenRouter, and cost roughly 6× less per call — while inference prices have fallen 50× in 36 months. However, open lags in operational tooling: only 51% of open-model teams reach production vs. 63% for closed, with the gap concentrated in deployment, compliance, and maintenance. The emerging battleground is the "agentic harness" above the model, where frontier labs are integrating scaffolding with weights to rebuild lock-in.