Meta Releases Muse Ember, a 10-Cent Lightweight Model Built for Agent Volume

Meta Superintelligence Lab released a new model called Muse Ember on Monday. According to TechCrunch, it is the lab's lightest and most cost-efficient model to date, with a deliberately narrow pitch: it is not competing to be the strongest reasoner, but to do the part of an agent's job that repeats all day.

Meta's release is specific about the workloads it has in mind: complex reasoning execution, plus the tool calls, coding subtasks, routing, extraction and long document parsing that agents perform over and over. These are high-frequency, low-novelty operations that also happen to be where inference bills spiral.

Price is the headline. Ember costs 10 cents per million input tokens, well below frontier competitors such as Anthropic's Haiku and OpenAI's Luna, and in the same band as open-source rivals like DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta leans on cost-per-task rather than raw price, using Artificial Analysis's composite benchmark score to argue that Ember is both cheaper and smarter than its competitors.

In practice Ember is built to be paired, not deployed alone. It can form an agentic harness with Meta's larger Muse Wonder, where the heavy model sets the rules for handling a given issue and the light model applies those rules across tens of thousands of cases. Meta sums it up in the announcement as "Muse Wonder plans, Muse Ember handles the volume." Customer support is the obvious example.

Because that kind of work arrives in messy formats, Meta trained Ember across text, images and audio. The model can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without a separate model behind it. For a model whose selling point is throughput, that removes a handoff from the pipeline.

Ember is the fourth model under the Muse brand. It follows the mid-weight Spark in April, Spark's open-source twin Glimmer in August, and the frontier-class Wonder in October. The consumer flagship Muse Agent launched in September and has been adding features such as Spaces and video chat this month, giving the family an unusually dense release cadence.

The real story is that price competition among lightweight models is shifting from cost per token to cost per task, and Ember is built squarely for that shift. Whether the two-model arrangement actually delivers the advertised economics under production agent traffic will depend on independent reruns and customer adoption, not on composite benchmark scores alone.