Meta has released Muse Ember, a lightweight model built for agent execution, and the headline number is its price: 10 cents per million input tokens.
A token is a small chunk of text, roughly a word fragment, and input tokens are the ones you send to the model. At that rate, Ember is significantly cheaper than frontier competitors like Haiku and Luna, according to coverage of the launch. That is the one comparison on offer, and it covers input pricing only.
The focus on agents explains why that line matters. An agent is software that works through a task in many steps, calling a model repeatedly along the way. Each call typically resends instructions, tool descriptions and earlier results, so input tokens tend to pile up faster than the text the model writes back. A low input price is therefore the right number to lead with for this kind of work.
It is not the whole bill, though. Total cost depends on output pricing, how many steps a task takes, and how often the model finishes the job without a retry. A cheap model that needs three attempts can cost more than a pricier one that needs one.
The pairing of "lightweight" with "agent execution" also suggests a division of labor: a smaller model carrying out steps, with larger ones reserved for harder reasoning. Whether Ember holds up on that job is a question for real workloads, not a price list.