Meta's Superintelligence Lab has introduced a new artificial intelligence model called Muse Ember, described as the lab's smallest and most economical offering to date. The model was made available on Monday, with Meta positioning it as a purpose-built tool for the kind of repetitive, high-volume tasks that AI agents handle continuously in production environments.

According to Meta's release documentation, Ember was built specifically for "complex reasoning execution and the work agents repeat all day," covering functions such as tool calls, coding subtasks, routing, data extraction, and long document parsing. The framing makes clear that this is not a general-purpose frontier model but rather a workhorse designed to operate efficiently at scale.

Competitive Pricing at the Core of the Pitch

Meta placed particular emphasis on Ember's cost-per-task advantage over rival lightweight models. The company cited benchmark data from Artificial Analysis, a composite scoring platform that evaluates AI models across multiple performance dimensions, to support its claim that Ember delivers better value than comparable offerings from competing labs.

Priced at 10 cents per million input tokens, Ember undercuts proprietary competitors including Anthropic's Haiku and OpenAI's Luna on cost while also matching the price points of open-source alternatives such as DeepSeek V4.1 Flash and GLM 5.3 Flash. That dual competitiveness - against both commercial and open-source rivals - is a notable positioning choice, suggesting Meta is targeting developers who might otherwise reach for free or low-cost community models.

Built to Work Alongside Larger Models

The practical architecture Meta envisions for Ember is one of collaboration rather than standalone deployment. Ember is intended to pair with heavier models - specifically Meta's own Muse Wonder - inside what the company refers to as an agentic harness. In this setup, a larger model handles high-level planning and rule-setting, while Ember takes on the downstream execution across large volumes of cases.

Meta illustrated this dynamic with a customer-support scenario: a more capable model might determine how a particular type of issue should be handled, then delegate the actual application of those rules across tens of thousands of individual interactions to Ember. As Meta put it in its announcement, "Muse Wonder plans, Muse Ember handles the volume." This division of labor is designed to keep costs manageable without sacrificing the quality of orchestration at the top of the pipeline.

Multimodal Capabilities Expand Its Utility

With high-volume, agent-driven deployments in mind, Meta also trained Ember to handle more than just text. The model supports a range of input types including images and audio, making it more flexible than its compact size and price might initially suggest. According to Meta's announcement, this means "agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model."

That multimodal capability is significant in practical terms. Rather than requiring developers to route different input types through separate specialized models - adding complexity and cost - Ember can process all of them natively. For enterprise deployments where agents routinely encounter mixed-media data, this reduces architectural overhead considerably.

The Fourth Model in Meta's Muse Lineup

Ember represents the fourth model released under Meta's Muse brand since the lineup began taking shape this year. The series launched with Spark, a mid-weight model released in April, followed by Glimmer, an open-source counterpart to Spark that arrived in August. Wonder, the flagship frontier model in the Muse family, was released in October and serves as the high-end anchor of the portfolio.

Alongside these model releases, Meta has also been building out its consumer-facing AI presence through the Muse Agent product, which launched in September and has reportedly performed beyond expectations in terms of user adoption. New features have continued to roll out at a steady pace, with additions such as Spaces and video chat functionality introduced earlier this month, pointing to an ongoing effort to expand the product's capabilities as its user base grows.

Taken together, the Muse lineup now spans a full range of use cases - from lightweight agent execution at the Ember end of the spectrum to heavy-duty reasoning and planning at the Wonder end - giving developers and enterprise customers a tiered set of options within a single product family.