Meta Ships Muse Ember, a 10-Cent Model Built for Agent Grunt Work

Meta's Superintelligence Lab on Monday released Muse Ember, its lightest and most cost-efficient AI model to date, priced at 10 cents per million input tokens. That price undercuts frontier rivals such as Anthropic's Haiku and OpenAI's Luna, and puts Ember in direct competition with open-source alternatives like DeepSeek V4.1 Flash and GLM 5.3 Flash.

Meta positioned the model squarely around cost-per-task. Using Artificial Analysis's composite benchmark score, the company argued Ember is simultaneously cheaper and smarter than Haiku and Luna. That is a vendor citing a third-party benchmark in its own launch materials — a claim worth watching once independent testers publish their own results.

What is Ember actually for?

Not frontier reasoning. Meta designed Ember for what it calls "complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." In plain terms, the model handles the high-volume, mechanical steps inside AI agents rather than the strategic layer.

The intended architecture is a two-model pairing. A heavier model — Meta's own Muse Wonder — sets the rules for a task, then Ember applies those rules across tens of thousands of cases. Meta's announcement puts it bluntly: "Muse Wonder plans, Muse Ember handles the volume." A customer-support agent is the template case: Wonder decides how to handle a category of complaint, Ember executes the same procedure at scale.

How flexible is the input handling?

More than its price suggests. Meta trained Ember across text, images, and audio, so a single deployment can absorb mixed inputs. "Agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model," the announcement states.

That matters for the economics. Running separate vision and speech models alongside a text model multiplies infrastructure cost. A single cheap model that handles all three removes a layer of plumbing — though Meta's own claims about cross-media accuracy remain unverified by outside benchmarks at launch.

Where does Ember sit in Meta's lineup?

Ember is the fourth model under the Muse brand, and the cadence is fast:

• Muse Spark — mid-weight model, released in April
• Glimmer — Spark's open-source twin, released in August
• Muse Agent — flagship consumer AI product, released in September
• Muse Wonder — frontier model, released in October

The consumer-facing Muse Agent hit what the source describes as unexpected success after its September launch, and Meta rolled out new features — Spaces and video chat — earlier this month. Ember now gives the agent stack a cheap execution layer underneath it.

What is the business logic here?

The pricing tells the story. At 10 cents per million input tokens, Ember costs a fraction of frontier models, which makes mass deployment of agents financially plausible where it previously was not. A support operation parsing tens of thousands of documents per day cannot do so on frontier-model pricing; it can on Ember's.

The open-source competition complicates the margin picture, though. DeepSeek V4.1 Flash and GLM 5.3 Flash already sit at comparable cost, so Meta is fighting on benchmark performance and its integrated Wonder pairing rather than on price alone. Meta did not disclose Ember's training costs, licensing terms, or whether an open-source variant — a Glimmer-style twin — will follow.

The takeaway: Meta is betting that the agent market's money sits in volume execution, not headline intelligence — and it has priced its newest model to win that specific fight.