Skip to main content
The Quantum Dispatch
Back to Home
Cover illustration for Meta Muse Ember: What a $0.10 Agent Model Means for Builders

Meta Muse Ember: What a $0.10 Agent Model Means for Builders

Meta's Muse Ember costs $0.10 per million input tokens and targets agent subtasks. Here's how it compares with Claude Haiku 5.5 and GPT-6 Luna.

Dr. Nova Chen
Dr. Nova Chen★Nov 13, 2026★4 min read

Not every step an AI agent takes needs a frontier-sized brain. Most of an agent's day is small, repetitive work: calling a tool, reading a file, routing a request, pulling a field out of a long document. On November 9, Meta Superintelligence Labs released Muse Ember, its smallest and most cost-efficient model, built specifically for that high-volume layer of agent work, and the pricing is aggressive enough to change how developers budget agent pipelines.

  • Price: $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens
  • Context: 1M-token context window with up to 512K output tokens; native text, image, audio and video understanding
  • Availability: Live now in Muse Code and the Meta Model API as muse-ember-1.0, with Responses API support, and on Google Cloud and Oracle Cloud
  • Headline result: Meta cites an Artificial Analysis Intelligence Index score of 36 at about $0.14 per task, matching Claude Haiku 5.5's best score at roughly a third of the cost

What Is Muse Ember Built For?

Meta frames Muse Ember as the execution layer of its model family. Muse Wonder handles the hardest reasoning, Muse Spark covers coding and agentic workflows, and Muse Ember takes the repeatable tasks an agent fires off hundreds of times per session. In Meta's own words, Wonder plans and Ember handles the volume. It is also the first model Meta says was trained using Muse Wonder Full, which fits the company's earlier promise to use its largest models to build smaller specialized ones. We covered the top of that family when Muse Wonder added active memory.

The 1M-token window matters more than it first appears for a small model. It means a single call can take in a full codebase or a large document set, so an agent does not need to chunk everything into tiny pieces before a cheap model can help.

How Does Muse Ember Compare With Claude Haiku 5.5 and GPT-6 Luna?

All of the comparison numbers below come from Meta's announcement, which mixes Artificial Analysis results with Meta's internal evaluations, so treat them as the vendor's claims until more independent testing lands. Interesting Engineering's report on the launch relays the same figures.

  • AutomationBench-AA (business workflows): Muse Ember 37.8%, GPT-6 Luna 20.2%, Claude Haiku 5.5 13.9%
  • Terminal-Bench 4.0 (agentic terminal coding): Muse Ember 40.9%, Haiku 5.5 39.2%, Luna 16.4%, each in its best published harness
  • FrontierCode 2.0-AA (mergeable engineering work): Haiku 5.5 leads at 34.9%, with Muse Ember at 27.2%
  • OSWorld 2.1 (computer use, partial credit): Muse Ember 69.5%, Haiku 5.5 68.9%, Luna 50.5%

The honest read is that Muse Ember is not a clean sweep. Haiku 5.5 still leads on the hardest real-engineering coding benchmark. Where Ember stands out is cost per unit of capability. Against Meta's own Muse Spark 1.3, Artificial Analysis estimates an Intelligence Index task costs $0.14 on Ember versus $2.39 on Spark, for a one-point difference in score (36 versus 37).

Why Cheap Subagent Models Matter

Agent frameworks increasingly split work between an expensive planner and cheaper workers. That is the same idea behind model routing between open and frontier models, and Muse Ember is built squarely for the worker role. Early partners point at exactly that use. Cline's founder described it as purpose-built for subagent work like coding and document analysis. Cloudflare's CTO said it ran up to 30% faster at the tail than the comparable model it tested. You.com reported a p95 latency of 2.2 seconds on its coding tasks with 36 to 48% fewer output tokens than leading open-weight models.

For teams running thousands of agent steps a day, a lower token bill and faster turnaround compound quickly. A pipeline that hands routine steps to a small model and saves the premium model for planning can cut costs without giving up much quality on the work that matters.

Safety Testing for a Small Model

Meta says Muse Ember went through the same core safety training and testing as its larger models under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and stayed below the framework's high-risk thresholds. The company also tested agent behavior: respecting user intent, handling sensitive data, resisting instructions hidden in web pages or tool outputs, and asking for human input before high-impact actions. That last group of checks is exactly what you want from a model designed to run unattended inside automated workflows.

Meta says even more efficient models are in training. For now, Muse Ember gives developers a strong, low-cost new option for the busy middle of the agent stack. For more model launches, browse our AI coverage.

Sources: Meta AI: Introducing Muse Ember — November 9, 2026; Interesting Engineering — November 9, 2026; VKTR — November 9, 2026.

More AI Stories