A lightweight model built for agents
Meta has released Muse Ember, a new lightweight model aimed at executing tasks for AI agents. Its main pitch is lowering costs in scenarios where agents call a model repeatedly in the background.
Price competition
According to Meta, Muse Ember costs 10 cents per million input tokens. The company compares this with frontier models such as Haiku and Luna, highlighting that Ember is significantly cheaper.
What it means
In agent-based applications, a single task can require many consecutive model calls, which quickly drives up token costs. By positioning a lightweight, low-cost model at this layer, Meta aims to let developers run agent products with lower operating costs.
