Meta Releases Muse Ember for Agent Work

Meta has released Muse Ember, a lightweight model designed for agent execution at a much lower price than frontier competitors. At 10 cents per million input tokens, Ember targets workloads where agents call tools, follow instructions and complete routine tasks repeatedly. The economics of agentic AI depend heavily on inference cost because agents may make many calls per job. A single task can involve planning, retrieving information, using software tools, checking results and trying again, which quickly multiplies token usage. A cheaper model can therefore make always-on automation more viable for startups and large teams alike. Muse Ember appears aimed at the practical middle ground between powerful frontier systems and low-cost utility models. It may not need to solve the hardest reasoning problems if it can reliably handle repetitive agent steps such as routing requests, extracting details, updating records and coordinating tool use. For businesses, that could open the door to broader automation in customer support, back-office operations, research assistance and internal workflows where cost per action is a major constraint. The release also shows Meta positioning itself in the agent economy, where success depends not only on model quality but also on price, latency and integration. Cheaper inference could encourage developers to experiment with multi-step agents that were previously too expensive to run at scale. The trade-off will be reliability: low-cost agents still need to follow instructions, avoid cascading errors and know when to escalate. If Ember performs well within those limits, it could help shift agentic AI from pilot projects toward more routine deployment.