Free guide

Meta Muse Ember: Let a Big Model Plan and a Cheap One Do the Work

Meta's new small model is built for the work agents repeat all day. Here's the planner and worker setup Meta used to port 16 code files to Rust for about $4, how to call Ember today, and the catches.

By Beau Johnson·
The Short this guide goes with: Meta Muse Ember: Let a Big Model Plan and a Cheap One Do the Work

Quick answer

Muse Ember is Meta's smallest and most efficient model, released November 9, 2026 on the Meta Model API (model ID muse-ember-1.0) and in Muse Code. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens. Meta's advice is to let a bigger model like Muse Wonder plan and review, and hand the repeated, bounded work to Ember.

The short version

  • Ember is built for tool calls, coding subtasks, routing, extraction and long documents, with a 1M-token context window and text, image, audio and video input.
  • In Meta's own test, Muse Wonder planned and 16 Ember agents ported 16 Pygments lexers from Python to Rust. Ember read 158M input tokens to Wonder's 4.4M, and the run cost about $3.96.
  • Per Meta, Ember scores 36 on the Artificial Analysis Intelligence Index, the same published score as Claude Haiku 5.5, at about a third of the cost per task. Haiku 5.5 still scores higher on FrontierCode 2.0.

How it actually works

Don't pay the head chef to peel potatoes

Most of an agent's tokens go to routine steps: reading files, searching, small edits, checks. If one big model does all of it, every one of those steps is billed at the big model's price. Ember is Meta's answer for that pile of work.

The big model writes the tickets

Meta's suggested split: a bigger model (Muse Wonder or Muse Spark) plans the change, Ember works each bounded subtask, and the bigger model reviews the combined result alongside a developer. The handoff carries the task, the relevant context, the files the worker may change, the acceptance checks and the result to return.

What the Rust port showed

In the Python to Rust port, Muse Wonder never opened the source. A worker mapped the code, Wonder wrote one brief per lexer, and 16 Ember agents each ported one lexer in its own copy of the repo. All 16 matched Python on 383 of 383 test inputs and 62 of 62 held-out files, and the Rust versions ran 3.0 to 7.3 times faster. Cost: $2.12 for Wonder, $1.84 for Ember. Meta notes results vary from run to run.

Cheap, not weak, but not the top coder

On Meta's own table Ember leads Haiku 5.5 and GPT-6 Luna on Terminal-Bench 4.0 (40.9 vs 39.2 and 16.4) and ties Haiku on the Artificial Analysis Intelligence Index (36). Haiku 5.5 still wins FrontierCode 2.0 (34.9 vs 27.2). Several of the other benchmarks in Meta's table are Meta's internal tests.

How to set it up

  1. 1.Get a Meta Model API key

    Sign in at dev.meta.ai, create an API key and keep it in an environment variable. The API is OpenAI SDK compatible, so you only need the openai package.

    pip install openai
    export MODEL_API_KEY="LLM|..."
  2. 2.Call Muse Ember

    Point the OpenAI client at Meta's base URL and pass the key yourself (the SDK won't read MODEL_API_KEY on its own). Ember supports the Responses API.

    import os
    from openai import OpenAI
    
    client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
    resp = client.responses.create(
        model="muse-ember-1.0",
        input="Extract the invoice number, issue date and total from this text: ...",
    )
    print(resp.output_text)
  3. 3.Or use it inside Muse Code

    Ember is included in every Muse Code plan ($5, $15, $50 or $200 a month). Install Muse Code with Meta's installer for Mac and Windows.

    curl -fsSL https://dev.meta.ai/install.sh | bash
  4. 4.Make Ember the worker and a big model the planner

    Meta's cookbook defines Ember as a subagent in an agents.json file and starts the main session on Muse Wonder. Give each worker a full brief: the files it owns, what to do and the commands that check the result.

    # agents.json
    {
      "worker": {
        "description": "Does one well-scoped task given in full in its input.",
        "model": "muse-ember-1.0",
        "effort": "max",
        "prompt": "Edit only the files your brief names, run its checks, then report."
      }
    }
    
    muse --model muse-wonder-1.0 --reasoning-effort high --agents "$(cat agents.json)"
  5. 5.Send the hard calls back up

    When a worker hits a broader dependency or an unclear requirement, Meta's advice is to send the finding back to the planner before continuing, and to have the bigger model review the combined changes with a developer.

The fine print

  • Price: $0.10 input, $0.004 cached input and $0.50 output per million tokens for prompts up to 200K tokens; $0.30, $0.012 and $1.50 above 200K.
  • 1M-token context window and up to 512K output tokens.
  • Ember is a hosted model: Meta offers it through the Meta Model API, Muse Code, Google Cloud and Oracle Cloud, with no download listed on its pages.
  • The $3.96 Rust port is Meta's own demo, and Meta says results vary from run to run.
  • Several benchmarks on Meta's page are Meta's internal tests; the Artificial Analysis numbers are independent.

FAQ

How much does Muse Ember cost?

$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens, with cached input at $0.004. Prompts over 200K tokens cost $0.30 in and $1.50 out.

What is Muse Ember good at?

Meta built it for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. It reads text, images, audio and video.

Is Muse Ember better than Claude Haiku 5.5?

It depends on the test. Both score 36 on the Artificial Analysis Intelligence Index, and Meta says Ember gets there at about a third of the cost per task. Haiku 5.5 still scores higher on FrontierCode 2.0, 34.9 to 27.2.

Is Muse Ember open source?

Not as far as Meta's pages show. Meta offers Ember as a hosted model through the Meta Model API, Muse Code, Google Cloud and Oracle Cloud, and lists no download.

Should Muse Ember be my main coding model?

Meta suggests pairing it: a bigger model like Muse Wonder or Muse Spark plans and reviews, and Ember handles the bounded, repeated tasks.

Sources

Facts checked 2026-11-10 against Meta's research blog, the dev.meta.ai Muse Ember build guide and pricing table, and Meta's launch posts on X.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides