What happened
Meta introduced Muse Ember on November 9 as a smaller model intended for the repetitive steps that make software agents expensive: calling tools, routing requests, extracting information, reading long documents and handling coding subtasks. The company offers the proprietary model through its Model API. Meta lists base prices of ten cents per million input tokens and fifty cents per million output tokens, with different pricing conditions for very long prompts. Its stated goal is not simply to win a general reasoning contest, but to lower the cost of executing many individual operations inside a larger workflow. VentureBeat independently reported the release and examined figures from Artificial Analysis, an outside evaluation organization. At maximum reasoning effort, the published comparison gave Ember an Intelligence Index score rounded to 36, alongside Claude Haiku 5.5 at the same rounded score.
Why it matters
The economic distinction matters because an agent can make dozens of model calls while completing one user request. A modest saving per call may become substantial at enterprise scale. The published evaluation estimated roughly fourteen cents per benchmark task for Ember versus about forty-one cents for the compared Haiku setting. That is a cost comparison under specified evaluation settings, not a guaranteed threefold reduction in any company's cloud bill. Model choice also depends on response quality, latency, context length, security controls and the number of retries a task requires. A cheaper model that makes more mistakes can cost more across a completed workflow.
What to watch next
Developers should compare Ember against existing small models on their own tasks, including tool reliability, document extraction accuracy and the cost of successful completed jobs rather than token price alone. Independent reproduction across multiple applications will help establish whether the launch measurements generalize. Teams should also test the impact of larger prompts and different reasoning settings, because pricing and performance can change substantially with workload design. The strongest evidence will come from repeatable production-like tests, not a single leaderboard position.