The economics of agent plumbing
Meta Superintelligence Lab released Muse Ember on Monday, and TechCrunch's coverage makes the pricing point immediately clear: this is a model aimed at cost per task, not conversational polish. The target is the unglamorous back office of agentic systems — tool calls, coding subtasks, routing, extraction, and long document parsing — the kind of work that repeats thousands of times a day.
The headline number is 10 cents per million input tokens. That undercuts Anthropic's Haiku and OpenAI's Luna, and it lands in the same cost band as open-source challengers DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta leans on Artificial Analysis' composite benchmark to argue Ember is both cheaper and smarter than those smaller frontier rivals. The comparison is self-selected, but the price alone changes the math for high-volume agent pipelines.
A multimodal workhorse under a planner
Ember isn't meant to think deeply on its own. Meta positions it as the execution layer beneath a larger model like Muse Wonder. In the customer-support example Meta gives, a heavier model establishes how a class of issue should be handled, then Ember applies those rules across tens of thousands of cases. The company's shorthand: “Muse Wonder plans, Muse Ember handles the volume.”
That division of labor explains Ember's unusual training emphasis. Meta built it to process text, images, and audio/video in a single model, so agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without bolting on a separate vision or speech module. For pipeline operators, each additional model is another latency hop and another failure point, so consolidating modalities matters.
Where Ember sits in the Muse stack
Ember is the fourth release under the Muse brand in a rapid cadence:
- Spark, the mid-weight model, shipped in April
- Glimmer, Spark's open-source twin, arrived in August
- Wonder, the frontier model, followed in October
- Ember now closes the year in November
The consumer-facing Muse Agent launched in September to what Meta calls unexpected success, with Spaces and video chat rolling out earlier this month. Ember's true payoff depends on agent traffic continuing to grow; a model priced for volume only becomes strategic if the volume actually arrives.
Meta Superintelligence Lab released a new AI model on Monday called Muse Ember, the lab’s lightest and most cost-efficient AI model yet. In its release, Meta said Ember was “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”
Meta particularly emphasized Ember’s cost-per-task, using Artificial Analysis’s composite benchmark score to show Ember as cheaper and smarter than competitors like Anthropic’s Haiku and OpenAI’s Luna.
At 10 cents per million input tokens, Ember is significantly cheaper than frontier competitors like Haiku and Luna. Not only that, it’s cost-competitive with open-source competitors like DeepSeek V4.1 Flash and GLM 5.3 Flash.