Meta releases Muse Ember, a new lightweight model for agent execution

Meta Superintelligence Labs released Muse Ember, its lightest and most cost-efficient model so far. Meta pitches it squarely at the repetitive work agents grind through all day: tool calls, coding subtasks, routing, extraction and long document parsing. The headline number is 10 cents per million input tokens, which TechCrunch notes undercuts frontier small models like Anthropic's Haiku and OpenAI's Luna and sits in the same range as open models like DeepSeek V4.1 Flash and GLM 5.3 Flash.

That matters because agent economics are volume economics. A planner model may make one expensive decision, but the follow-through is thousands of cheap calls. Whoever owns that bottom layer at a good price owns a lot of the bill.

We think this is the right kind of model to be building, and we're glad to see Meta compete on cost-per-task instead of leaderboard vanity. Most real agent deployments fail on economics before they fail on intelligence. If a support agent has to read tens of thousands of tickets, a model that is good enough and dramatically cheaper is the thing that makes the project exist at all.

Meta's framing is a split brain: Muse Wonder plans, Ember executes. That's a sensible architecture and it mirrors how teams are already wiring agents. The heavier model sets the rules, the lighter one applies them at scale. The multimodal piece is the quiet win here. Meta says Ember can read screenshots, charts, scanned documents, voice memos and meeting recordings without a separate model, which removes a lot of glue code and a lot of failure points.

Now the caution. The benchmark story leans on Artificial Analysis's composite score as Meta presents it, and 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 the ten-thousandth call. I'd want independent testing on exactly the volume tasks Meta is advertising before calling Ember a default.

There's also a bigger trajectory worth naming. When execution gets this cheap, the temptation is to let agents do far more unsupervised work, because the marginal cost of one more pass rounds to zero. That's useful right up until an error pattern replicates across tens of thousands of cases before anyone notices. Cheap volume models need cheap volume checks, and the industry is still better at the first than the second.

Still, the direction is good. A four-model lineup in roughly seven months, from Spark to Glimmer to Wonder to Ember, shows Meta building a full stack rather than one flagship. Competition at the bottom of the price curve pushes everyone's agent costs down, and that benefits small teams most.

Ember is the fourth model under Meta's Muse brand. It follows the mid-weight Spark in April, Spark's open source twin Glimmer in August, and the Wonder frontier model in October. Meta's consumer Muse Agent launched in September and, per TechCrunch, saw unexpected success, with Spaces and video chat features rolling out earlier this month.