Muse Ember: what it is, pricing and how to use it

Muse Ember is Meta's smallest Muse model, released November 9, 2026, for the repetitive work AI agents do all day. It costs $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API, and you can call it today as muse-ember-1.0.

What Meta posted on X

Mark Zuckerberg announced it himself. @finkd called it "our smallest model that delivers higher intelligence per dollar than others in its class," and aimed it at token-heavy chores: tool calls, coding subtasks, routing, extraction and long document parsing.

@AIatMeta added the details, describing a model "built for complex reasoning execution and everyday agent work." The same post claims results comparable to Claude Haiku 5.5 at about a third of the cost on the Artificial Analysis Intelligence Index. Ember arrives four weeks after Muse Wonder, and Meta's announcement says it is the first model trained using Muse Wonder Full.

Muse Ember pricing

These are the per-million-token rates from Meta's pricing page:

Watch the 200K line. Once a request goes over it, the whole request is billed at the higher rates, and cached tokens count toward the limit. A 250K request costs three times as much per token as a 190K one, so the 1M context window is there when you need it, but splitting long jobs saves real money.

By my math, sorting 1,000 customer emails at 2,000 tokens in and 200 tokens out comes to about $0.30 in total. That leaves out any reasoning the model does, and effort defaults to max, so run a sample and read the usage figures before you budget. A caption in Meta's announcement also still calls Ember's costs "estimates at proposed pricing," so check the pricing page again before you commit.

How to use Muse Ember

Ember is available in Muse Code and on the Meta Model API. AI Release Tracker lists the weights as unpublished, so don't plan on running it yourself. A minimal API call, based on Meta's model docs, looks like this:

curl -X POST "https://api.meta.ai/v1/responses" \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "muse-ember-1.0",
    "input": "Sort this customer message as billing, scheduling or complaint, and give a one-sentence reason: My invoice shows the wrong service date."
  }'

The docs say Ember takes the same request body as Muse Spark, so moving over is mostly a model ID change. It works on the Responses, Chat Completions and Messages endpoints, takes text, images, video, audio and PDFs, and accepts reasoning effort from minimal to max. Limits are 3,000 requests and 4 million tokens per minute, per team. OpenRouter also lists it as meta/muse-ember.

Three jobs are worth testing first:

• Routing: sort incoming leads or support tickets into your own categories.
• Extraction: pull the date, amount and vendor from a stack of PDF invoices.
• Subtasks: let a larger model plan and hand the steps to Ember, the split Meta's docs recommend.

How good is it?

AlphaSignal's write-up reports a score of 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5, at an estimated $0.14 per test task against $0.41 for Haiku. Ember generated about a third as many output tokens per task, which explains most of that gap. Haiku's launch score of 43 came from an older version of the index, so don't compare the two numbers.

On Artificial Analysis's AA-Omniscience knowledge test, AlphaSignal reports 21.1% accuracy, largely because Ember declined 64% of the questions. In Meta's own table, Haiku beats it on FrontierCode 2.0, 34.9% to 27.2%. Several of Ember's wins come from evaluations Meta built itself, and its Terminal-Bench comparison ran each model in a different harness. Meta's docs send hard single-shot coding, knowledge-heavy questions and visual coding to Muse Spark or Muse Wonder.

I'd use it for work where you hand it the material, like sorting and extraction, and keep it away from answering customers from memory. A model that declines is safer in an automation than one that guesses, but only if a declined answer goes somewhere, like a person's queue.

What this means for businesses using AI

When Meta opened its API in July, Muse Spark's $1.25 input price looked aggressive. Ember's input rate is less than a tenth of that, which makes high-volume steps like lead sorting and invoice parsing very cheap to run. You still have to decide which steps are safe to automate and what happens when the model declines, and New Face Design's free process audit can help you work that out.