Meta releases low-cost Muse Ember for AI agent workloads

Meta Superintelligence Lab announced Ember on Monday, highlighting its cost per task and performance on Artificial Analysis’s composite benchmark. Meta said the model compares favorably with Anthropic’s Haiku and OpenAI’s Luna. Input tokens cost 10 cents per million, a price the company said is also competitive with open-source models DeepSeek V4.1 Flash and GLM 5.3 Flash.

Meta expects Ember to work in an agentic system alongside a larger model such as Muse Wonder. In a customer-support example, Wonder could set rules for handling an issue, while Ember applies them across tens of thousands of cases. The company summarized the division of labor this way: “Muse Wonder plans, Muse Ember handles the volume.”

Ember was trained to process text, images and audio. Meta said agents can use it to read screenshots, charts and scanned documents, or work with voice memos, meeting recordings and video without a separate model.

Ember is the fourth model in Meta’s Muse line. It follows the mid-weight Spark, released in April; the open-source model Glimmer, released in August; and the frontier model Wonder, released in October. Meta’s Muse Agent consumer product launched in September and has since added features including Spaces and video chat earlier this month.