Meta launches Muse Ember, a lightweight agent model
Original: Meta releases Muse Ember, a new lightweight model for agent execution
Why This Matters
Cheap, capable execution models are the missing piece for scaling real-world AI agents.
Meta Superintelligence Lab released Muse Ember on Nov. 9, 2026 — a multimodal AI model priced at $0.10 per million input tokens, designed for high-volume agentic tasks like tool calls, routing, and document parsing.
Meta Superintelligence Lab's fourth Muse-branded model, Ember, targets the repetitive execution layer of AI agent workflows rather than top-level reasoning. At 10 cents per million input tokens, it undercuts Anthropic's Haiku and OpenAI's Luna on cost while matching or beating them on Artificial Analysis's composite benchmark. Meta also claims it's cost-competitive with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash.
The intended architecture is a two-tier setup: a heavier model like Muse Wonder handles planning and rule-setting, while Ember grinds through the volume — think tens of thousands of customer-support cases once a larger model has defined the approach. "Muse Wonder plans, Muse Ember handles the volume," Meta said in the announcement.
Ember is also multimodal out of the box, handling text, images, and audio — screenshots, charts, scanned documents, voice memos, and video — without requiring a separate specialist model. Ember follows Muse Spark (April), open-source Muse Glimmer (August), and frontier model Muse Wonder (October). The consumer-facing Muse Agent, launched in September, has seen strong early adoption.