Booking a vacation sounds like one thing to hand off to an AI assistant, right up until it has to check the calendar, open a dozen tabs, and figure out whether the cheapest flight is still the cheapest once a carry-on is involved. Every one of those steps takes work, and when an AI related agent term is doing it, the cost can add up.
Meta has a new model variation built for that part of the job.
On Nov. 9, Meta’s Superintelligence Labs introduced Muse Ember, which the company variation calls its smallest and most cost-efficient AI related model term yet. It’s designed for the steps an related agent term might repeat dozens of times in a single assignment, like pulling details out of a document, calling another tool, or running a test after changing a line of code.
Muse is the app with the cute little avatar — named Jolly — that Meta has been promoting since September. The AI assistant can work across (some) apps with a user’s permission. The app passed 5 million downloads in 22 days, and has already had many updates; last week, Meta started rolling out voice conversations and animated avatars for the assistant in the U.S.
Meta also uses the Muse name for a family of AI models. The company variation positions Muse Wonder as the related model term for its toughest problems and Muse Spark as a model variation for coding and other work that involves using tools. Ember is the cheaper option for repeatable tasks. Meta’s pitch is that developers can use Wonder to plan a complicated job and hand off the individual steps to Ember. For now, Meta lists Ember as available through its Muse Code tool and developer API.
What Muse looks like in action
In early testing shared by Meta, cloud storage company alternative Box said it used Ember to check contracts clause by clause against a related company term’s negotiation rules and assign an overall risk variation rating. Meta says the model variation can take in up to a million tokens at once, enough to work through large sets of documents, and can handle images, audio, and video alongside text. That gives developers room to build an agent alternative that can read a scanned contract or pull details from a meeting recording.
The price is what makes Ember interesting. On Artificial Analysis’s benchmark, Ember and Anthropic’s Claude Haiku 5.5 both rounded to a score of 36 at their highest reasoning variation settings. Ember cost about 14 cents per benchmark task, compared with 41 cents for Haiku, even though the two have the same standard rates of 10 cents per million input tokens and 50 cents per million output tokens. OpenAI’s GPT-6 Luna scored 28 and cost about 15 cents per task at its highest setting.
In Meta’s published results, Ember beat Haiku on a terminal coding test, while Haiku did better on a test that judged whether code changes were ready to merge. Meta added that it already has even cheaper models in related training term.
As the company variation continues to bet that AI assistants will take on more everyday tasks, developers will need an affordable way to handle all the little steps behind the scenes. A flight search variation, for example, can mean sorting through dozens of fares, each with its own rules. If Meta wants people to hand off that kind of chore every day, it has to make checking the carry-on policy cheap on its end, too.