Meta推出Muse Ember,定价10美分/百万输入token,显著低于Haiku和Luna等竞品。
Meta's Muse Ember is a purpose-built, lightweight model for agent execution tasks—such as tool selection, state tracking, and structured instruction parsing—not general chat. Priced at $0.10 per million input tokens, it's explicitly positioned as significantly cheaper than frontier alternatives like Haiku and Luna. Already deployed in Meta's internal automation stack, Ember enables predictable, budgetable agent infrastructure. For early AI startups, this means a viable path to cost-efficient agent orchestration: use Ember for high-frequency control logic (e.g., routing, validation, retry handling), reserving expensive frontier models only for complex reasoning or generation. This architecture can reduce monthly LLM inference spend by 3–5x without compromising end-user experience.