Meta Superintelligence Lab released Muse Ember on Monday, a lightweight AI model designed for high-volume agentic workloads at 10 cents per million input tokens—roughly one-tenth the cost of Anthropic's Haiku and OpenAI's Luna.
The model targets the economics of agent execution: tool calls, coding subtasks, routing, data extraction, and document parsing. Meta's architectural strategy pairs Ember with its more powerful Muse Wonder model in what amounts to a labor division—Wonder handles planning and decision logic; Ember executes across thousands of instances. The company described the division as "Muse Wonder plans, Muse Ember handles the volume."
This tiered approach optimizes for a specific cost structure: reserving expensive compute for infrequent, high-stakes decisions and routing repetitive work to cheap inference. In enterprise AI deployments, that math matters. If an agent makes 100 routing decisions and 10,000 execution steps per workflow, pricing becomes the dominant variable in unit economics.
Ember's multimodal training—images, audio, video, documents—eliminates the need for separate specialist models within an agentic pipeline, another cost lever for builders.
Ember is the fourth model in Meta's Muse line in under a year, following Muse Spark (April), Muse Glimmer open-source (August), and Muse Wonder (October). Meta's consumer Muse Agent, launched in September, now includes Spaces and video chat features.
The real competitive pressure lands not on pricing alone but on agent reliability and latency. Frontier models still command higher token costs for good reason—they handle exceptions and complex reasoning better. Meta's thesis: use them sparingly, use Ember for scale. Whether enterprise customers believe that tradeoff depends on error rates and the cost of agent hallucinations at volume.
