Meta launches Muse Ember, a lightweight model priced for high-volume agent work

Meta's lightest model targets agent execution

Meta's Superintelligence Lab released a new model on Monday called Muse Ember, which Meta describes as the lightest and most cost-efficient model the lab has shipped. According to TechCrunch, Ember is aimed squarely at agent execution: the routine, high-volume work that fills an agent's day, including tool calls, coding subtasks, routing, data extraction, and parsing long documents.

The distinction matters. Most model releases compete on reasoning benchmarks, but a deployed agent spends far more of its run on small repeated steps — reading a page, calling a tool, formatting a result — than on hard problems, and those steps are what Ember is built to carry.

Priced for volume

Ember costs 10 cents per million input tokens. TechCrunch reports that this significantly undercuts small frontier models such as Anthropic's Haiku and OpenAI's Luna, and puts Ember in cost competition with open-source options including DeepSeek V4.1 Flash and GLM 5.3 Flash.

Meta also used Artificial Analysis's composite benchmark score to argue that Ember is both cheaper and more capable than those rivals, according to TechCrunch — a framing that puts the emphasis on cost per completed task rather than raw intelligence.

Built to pair with Muse Wonder

Ember is not meant to run alone. According to TechCrunch, it is designed to work inside an agentic harness alongside a larger model such as Meta's Muse Wonder, with the big model doing the planning and Ember absorbing the day-to-day volume.

The customer-support example TechCrunch cites illustrates the split: a heavier model works out how to handle a given category of issue, then the cheaper model applies those rules across tens of thousands of individual cases. In that arrangement, the execution model consumes the overwhelming majority of tokens, so its price and reliability effectively determine the economics of the whole system.

One model for text, images, and audio

Because volume-heavy work arrives in messy formats, Meta trained Ember across multiple media types. According to the company, an agent running Ember can process screenshots, charts and scanned documents, and work from voice memos, meeting recordings and video, without adding a separate model for each input type. For document- or voice-heavy workflows, that removes a common source of integration work and per-modality cost.

The fourth model in the Muse family

Ember is the fourth release under Meta's Muse brand, TechCrunch reports. The mid-weight Spark arrived in April, its open-source twin Glimmer followed in August, and the frontier model Wonder landed in October. The consumer-facing Muse Agent launched in September and, per TechCrunch, met unexpected success, with features such as Spaces and video chat rolling out earlier this month.

Why it matters

Ember is a bet on where the money in AI actually flows. Frontier models attract the attention, but deployed agents are dominated by execution — the tool calls, extractions and parsing steps that repeat thousands of times a day. If most tokens in a production system pass through the cheap worker model, that model's pricing and reliability set the system's cost floor, not the planner's.

It is also a competitive signal. By pricing at 10 cents per million input tokens and benchmarking against both closed rivals and open-source ones, Meta is treating the small-model tier as a market worth winning outright. And the two-model pitch — a planner and a worker sold as a pair — nudges developers toward adopting Meta's stack end to end rather than mixing components from different vendors.