Meta Targets Agent Execution Layer With Low-Cost Muse Ember Release
AI-generated image · US National Wire
By pricing its lightest model to undercut frontier competitors, Meta is positioning Muse Ember as the high-volume engine for agentic workflows.
Meta Superintelligence Lab has released Muse Ember, a lightweight AI model specifically engineered for the repetitive, high-volume tasks associated with agent execution. According to reporting from TechCrunch and a technical blog post from Meta, the model is designed to handle tool calls, routing, extraction, coding subtasks, and the parsing of long documents.
From a deals and margins perspective, the most significant aspect of the release is Meta's aggressive pricing strategy. Through the Meta Model API, Muse Ember is priced at $0.10 per million input tokens and $0.50 per million output tokens. TechCrunch notes that this pricing makes Ember significantly cheaper than frontier competitors, specifically OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5. Meta further claims that Ember is cost-competitive with open-source alternatives, including GLM 5.3 Flash and DeepSeek V4.1 Flash.
Meta is positioning Ember not as a standalone intelligence, but as a component of an "agentic harness." The company's strategy involves pairing Ember with its frontier model, Muse Wonder. In this architecture, Wonder serves as the planning layer that establishes rules, while Ember handles the high-volume execution. Meta describes this relationship as: "Muse Wonder plans, Muse Ember handles the volume."
To support this high-throughput role, Meta trained Ember for multimodal flexibility, allowing agents to process voice memos, video, screenshots, and scanned documents without requiring separate models. The model also features a 1M-token context window and up to 512K output tokens, which Meta says allows it to ingest entire document sets or codebases in a single call.
Early adoption data provided by Meta includes several enterprise partners. Yashodha Bhavnani, VP of AI Products at Box, stated that Ember can score contracts against negotiation playbooks and screen candidates against briefs. Dane Knecht, CTO of Cloudflare, reported that the model ran up to 30% faster at the tail than comparable models in their testing. Additionally, Chak Pothina, Inference Lead at You.com, claimed Ember achieved a p95 latency of 2.2 seconds, noting this is between 4x and 13x faster than top open-weight models.
Muse Ember is the fourth release under the Muse brand, following the mid-weight Spark (April), the open-source Glimmer (August), and the frontier Wonder model (October). This follows the September launch of the Muse Agent consumer product.