Meta’s Muse Ember prices agent grunt work at 10 cents per million tokens

TechCrunch reports that Meta Superintelligence Labs has launched Muse Ember, the fourth model under the Muse brand and its lightest yet. Meta aims it at the repetitive work agents do all day: tool calls, coding subtasks, routing, data extraction and long-document parsing. The headline figure is 10 cents per million input tokens.

Ember completes a lineup built in roughly seven months: the mid-weight Spark in April, its open-source counterpart Glimmer, and the frontier Muse Wonder in October, which we covered in Meta releases its first frontier model. For comparison, Meta priced Wonder at $4 per million input tokens and $20 per million output tokens on its public API (Wonder pricing details).

A planning model might make one expensive decision, but carrying it out can take thousands of cheap calls. Whoever owns that bottom layer at a good price captures a large share of the bill. Most real agent deployments fail on cost before they fail on intelligence. If a support agent must read tens of thousands of tickets, a model that is good enough and dramatically cheaper is what makes the project viable at all. Competing on cost per task rather than leaderboard position is the right instinct.

Meta’s split-brain design, Wonder plans and Ember executes, mirrors how many teams already wire agents: a heavier model sets the rules and a lighter one applies them at scale. The multimodal piece is a quiet win. Reading screenshots, charts, scanned documents, voice memos and meeting recordings without a separate model removes glue code and failure points.

Meta’s performance story leans on Artificial Analysis’s composite score as Meta presents it. Composite scores flatten the differences that matter in production: how a model behaves on your tool schema, how often it silently drops a field during extraction, how it handles a malformed document. Cheap and smart on a chart is not the same as reliable on call number ten thousand. Independent testing on exactly the volume tasks Meta advertises should come before anyone makes Ember a default.

There is also a second-order risk. When execution is this cheap, it is tempting to let agents do far more unsupervised work because one more pass costs almost nothing. That holds until an error pattern repeats across tens of thousands of cases before anyone notices. Cheap volume models need cheap volume checks, and the industry is better at the first than the second.

Meta has been building a full stack rather than a single flagship, and using price aggressively: it previously offered Muse Spark discounts of roughly 95 percent to developers who share prompts and outputs for training (our coverage). The consumer Muse agent launched in September and, according to TechCrunch, has been an unexpected success. Price pressure at the bottom of the curve pushes everyone’s agent costs down, which helps small teams most. More in our Meta coverage and agents coverage.