Meta Superintelligence Labs has unveiled its new model, Muse Ember, the smallest AI model in the Muse family to date. This model was developed specifically to serve as the executive engine for digital agents (AI Agents), focusing on performing repetitive and specific tasks within workflows, such as writing simple code, extracting data from documents, and classifying customer requests, all at a very low operational cost compared to massive models.
Planning and Execution Strategy: Integration Between Models
Meta's new philosophy relies on dividing labor between different levels of artificial intelligence. In this system, large models like Muse Wonder or Muse Spark handle strategic planning and complex decision-making tasks. For example, the Wonder model can determine refund policies for a customer, while Muse Ember handles the execution of these rules across thousands of requests, verifying order logs, and routing cases that require human intervention to the relevant teams.
This separation between the "planning brain" and the "executing muscles" aims to reduce the overall bill for developers, as using a small model for routine tasks significantly lowers the cost per inference compared to using a massive model for every step.
Technical Capabilities and Multimodality
Muse Ember is not limited to text processing; it features multimodal capabilities, allowing it to receive inputs including images, audio, video, and PDF files, while producing text outputs. This makes it ideal for tasks such as extracting payment dates from scanned invoices or summarizing action items from audio meeting recordings.
- Context Window: Reaches up to one million tokens, allowing the model to examine massive sets of documents or codebases in a single request.
- Output Limit: The model can produce up to 524,000 tokens in a single response.
Cost and Performance in Benchmarks
Meta has introduced a competitive pricing structure via Meta Model API and Muse Code. The cost for one million input tokens starts at $0.10 (for prompts not exceeding 200,000 tokens), while the cost for one million output tokens is $0.50. These prices increase if the prompt exceeds the 200,000-token limit, reaching $0.30 for inputs and $1.50 for outputs.
Tests showed a notable superiority for Muse Ember in certain areas; in the AutomationBench-AA test, which measures multi-step processes, the model achieved a success rate of 37.8%, outperforming OpenAI's GPT-6 Luna model (20.2%) and Anthropic's Claude Haiku 5.5 model (13.9%). However, in the FrontierCode 2.0-AA programming test, the Claude Haiku 5.5 model outperformed it with 34.9% compared to 27.2% for Muse Ember.
What Changes Practically?
The launch of Muse Ember represents a shift in how AI applications are built; instead of relying on a single massive and expensive model to do everything, developers are moving toward a hybrid infrastructure. This approach solves the problem of high costs and slowness in executing routine tasks, opening the door for digital agents capable of working at scale and with economic efficiency that makes full automation of administrative and simple programming tasks a realistic and commercially viable endeavor.