Meta released Muse Ember on Monday, its smallest and cheapest model yet. The price is the news: 10 cents per million input tokens. That is well below near-frontier models such as Anthropic's Haiku or OpenAI's Luna, and competitive with cheap open variants like DeepSeek V4.1 Flash and GLM 5.3 Flash.
Ember is positioned not as a thinker but as a worker. Meta describes it as built for "complex reasoning execution and the work agents repeat all day": tool calls, coding subtasks, routing, extraction, parsing long documents. In an agent harness it is meant to pair with a larger model such as Muse Wonder — Wonder plans, Ember handles the volume. In customer support that means a heavy model sets the rules and a light one applies them across tens of thousands of cases.
To that end Ember is broader than the price suggests: text, images and audio without a separate model. Agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video.
Meta argues with cost per task rather than benchmark peaks, using Artificial Analysis's composite index. That is the more honest metric for agent economics — and also the easiest one to flatter, since it depends on the chosen task mix. Independent re-measurement on real coding and extraction work will show whether Ember keeps its price advantage or pays for it in retries.
Ember is the fourth model in the Muse family, after Spark in April, its open-source twin Glimmer in August, and the frontier model Wonder in October. The strategy is legible: whoever supplies the billions of cheap agent steps does not have to win the expensive planner.

