Input tokens cost $0.10 per million, outputs cost $0.50, and cache-hit inputs only cost $0.004. This is the price list for Meta’s new model, the Muse Ember, and what it’s most likely to be remembered for.
Ember is Meta's currently smallest and most efficient model, and its positioning is very narrow: high-frequency activities that need to be run repeatedly in the agent, such as adjusting tools, performing programming subtasks, request routing, information extraction, and parsing long documents.
A few key figures:
- The context is about 1 million tokens, and the maximum output at a time is about 512,000 tokens.
- Artificial Analysis comprehensive intelligence index is 36 points, which is the same as the highest reasoning level of Claude Haiku 5.5
- The cost of a single review task is about one-third that of Haiku 5.5
- The end-to-end business process test AutomationBench-AA scored 37.8 points, Haiku 5.5 scored 13.9, and GPT-6 Luna scored 20.2
It natively supports text, image, audio and video input, and screenshots, scans, and meeting recordings can be fed directly. Already available on Meta Model API, Google Cloud and Oracle Cloud.
I care more about the AutomationBench set of numbers. The comprehensive index is tied, which only means that the two are almost as smart; the business process test is more than twice as advanced, which means that Meta has been specially adjusted for the intelligent agent pipeline. The small model does chores in the intelligent body. The chores can be performed stably and cheaply, and are useful if they are a few points higher than the overall score.
The shortcomings are also obvious. Meta’s developer API has not been tested for large-scale production traffic, and many teams will not cut off core traffic just because it is two-thirds cheaper. It is safer to try it first in links such as routing and extraction where the cost of errors is low.
Meta AI has had its ups and downs on large models in the past two years, and Ember has at least found a clear position.