Muse Ember is the cheapest model Meta released on 2026-11-09: input costs just $0.10 per million tokens, the model ID is muse-ember-1.0, it is called through the Meta Model API, and it is also live in Muse Code. Cached input costs $0.004 per million tokens, and output costs $0.50 per million tokens.
What it can be used for
Muse Ember is built for the high-frequency, repetitive work agents do every day: tool calls, coding subtasks, routing, information extraction and parsing long documents. It has a 1 million token context window and outputs at most 512K tokens in a single call. It natively supports text, image, audio and video input, and can read screenshots, charts, scanned documents, voice memos and videos. The official figure is about $0.14 per task, roughly a third of Claude Haiku; it scores 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5's best score. In the division of labour, Muse Wonder plans the hardest problems while Ember handles execution at scale. Muse Ember, Muse Spark and Muse Wonder are now live on Google Cloud and Oracle Cloud.
Three things to check before calling it
First, how to call it: use the Responses API through the Meta Model API with the model ID muse-ember-1.0; it is also available in Muse Code, and the exact subscription tiers, free allowances and rate limits are as stated in the official documentation. Second, do not treat it as an all-purpose model: knowledge questions are a weak spot, and a third-party account of Artificial Analysis puts the refusal rate on trivia questions at about 64%, so check facts and general knowledge first. Third, the parameter count has not been published and the weights are not open source, so do not plan to deploy it the way you would an open-source model.
