Agentic AI
Meta launches Muse Ember, a cost-efficient AI model built for agent tasks at scale
The lightweight model excels at handling repetitive work like tool calls, coding, and document parsing—all while undercutting rivals on price.
1/2
TechCrunch — AI
Meta's Superintelligence Lab unveiled Muse Ember this week, a streamlined AI model engineered to power the high-volume, repetitive operations that autonomous agents handle daily. The system targets a sweet spot between affordability and capability, priced at just 10 cents per million input tokens—significantly undercutting established competitors.
Positioning and Performance#
Ember is positioned as the lightest in Meta's Muse lineup and sets itself apart through cost efficiency. According to Artificial Analysis's benchmark, the model delivers better performance per dollar than Anthropic's Haiku or OpenAI's Luna, and matches the economics of open-source alternatives like DeepSeek V4.1 Flash and GLM 5.3 Flash.

Built for Agent Workflows#
The model specializes in the core tasks agents repeat constantly: making tool calls, executing coding snippets, routing requests, extracting data, and processing lengthy documents. Meta designed Ember to pair with its more capable Muse Wonder model in a tiered architecture—Wonder handles strategic planning while Ember executes decisions across thousands of cases, as Meta summarizes it: "Muse Wonder plans, Muse Ember handles the volume."
This division of labor makes Ember ideal for scenarios like customer support systems, where a heavyweight model might define handling rules and an Ember instance applies them consistently across tens of thousands of incoming tickets.
Multimodal Capabilities#
Unlike many lightweight competitors, Ember can process text, images, and audio, giving it unusual flexibility for a cost-optimized model. Agents can work with screenshots, charts, scanned documents, voice memos, meeting recordings, and video—eliminating the need for separate specialized models.
Part of a Growing Portfolio#
Ember joins Meta's expanding Muse family: the mid-weight Spark (April), its open-source counterpart Glimmer (August), and the frontier-class Wonder (October). Meta's consumer-facing Muse Agent, , has already found , with features like and video chat rolling out this month.

