Meta introduced Muse Ember on November 9, a model it calls its smallest and most cost-efficient yet, built for the cheap, repetitive steps inside an AI agent's workload.
Ember targets work an agent might repeat dozens of times in one assignment: pulling details out of a document, calling another tool, or running a test after a code change. It sits in a Muse model family from Meta's Superintelligence Labs. Muse Wonder is pitched for the toughest problems, and Muse Spark for coding and tool use. Meta's idea is that developers let Wonder plan a complicated job, then hand the individual steps to Ember. For now, Ember is available through Meta's Muse Code tool and its developer API.
The model can take in up to a million tokens (the chunks of text a model reads) at once, and it handles images, audio and video alongside text. Box reported using Ember in early testing shared by Meta to check contracts clause by clause against a company's negotiation rules and assign an overall risk rating.
The pricing numbers are the interesting part. On the Artificial Analysis benchmark, Ember and Anthropic's Claude Haiku 5.5 both rounded to a score of 36 at their highest reasoning settings. Ember cost about 14 cents per task. Haiku cost 41 cents. OpenAI's GPT-6 Luna scored 28 at about 15 cents.
That gap is odd, because Ember and Haiku carry the same standard rates: 10 cents per million input tokens and 50 cents per million output tokens. One plausible reading is that Haiku uses more tokens per task at its top setting, and the bill follows the token count. That is an inference from the matching rates. It also shows why a price list tells you little about what an agent will cost to run.
Meta's own results are mixed. Ember beat Haiku on a terminal coding test, while Haiku did better on a test of whether code changes were ready to merge. Meta added that even cheaper models are in training.
The Muse assistant app passed 5 million downloads in 22 days, so Meta has an audience for agent chores. A flight search can mean sorting dozens of fares, each with its own rules. Running that every day only works if checking the carry-on policy costs Meta almost nothing.