Closed SourceMetaReleased 2026-11

Muse Ember

Meta's smallest, cheapest Muse model, built for the repeated steps inside agents

Context window

1M

Input / 1M tokens

$0.10

Output / 1M tokens

$0.50

Provider

Meta

Meta Model API, per 1M tokens for prompts up to 200K tokens: input $0.10, cached input $0.004, output $0.50. Prompts over 200K tokens cost $0.30 input, $0.012 cached and $1.50 output. Standard mode only, with no discounted Contributor mode. Max output 524,288 tokens. · Data verified 2026-11-11

Muse Ember (November 9, 2026) is the smallest and most efficient model in Meta's Muse family, built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta positions it as the execution layer under Muse Wonder or Muse Spark, which plan while Ember handles the volume, and says it is the first model trained using Muse Wonder Full. It costs $0.10/$0.50 per 1M input/output tokens, the same list price as Claude Haiku 5.5 and GPT-6 Luna, and Meta claims more intelligence per dollar than both. It takes text, image, video, audio and PDF input, with a 1M-token context and up to 512K output tokens.

How to access it

Meta Model API (model ID muse-ember-1.0, Standard mode only) through the Responses, Chat Completions and Messages endpoints, and in Muse Code on pay-as-you-go and every subscription plan. Also listed on OpenRouter as meta/muse-ember. Meta says Muse Ember is also available on Google Cloud and Oracle Cloud.

Strengths

  • ✓$0.10/$0.50 per 1M with cached input at $0.004
  • ✓Terminal-Bench 4.0 40.9% at max effort vs 39.2% for Claude Haiku 5.5 and 16.4% for GPT-6 Luna (Meta-reported)
  • ✓Native text, image, video, audio and PDF input
  • ✓1M-token context and up to 512K output tokens
  • ✓Accepts the same requests as Muse Spark across the Responses, Chat Completions and Messages APIs

Best for developers who...

Sub-agent work in coding agents: reading files, running tests, targeted fixesHigh-volume tool calling, routing and classificationExtraction across large collections of documents, images or records

When to choose it (and when not to)

Reach for Muse Ember when...

  • →You run many small agent steps (tool calls, routing, extraction) where cost per call adds up
  • →You want a cheap sub-agent under Muse Wonder or Muse Spark with the same request format
  • →You need audio or video input in a low-cost model

Look elsewhere if...

  • ✕Hard single-shot coding problems or knowledge-heavy questions, where Meta points to Muse Spark or Muse Wonder
  • ✕Prompts over 200K tokens when cost matters (rates triple)
  • ✕You need open weights (API only)

How to use it

  • ›Reasoning effort runs from minimal to max and defaults to max; lower it for routing and extraction steps
  • ›Use structured output for extraction and routing so later steps get schema-valid JSON
  • ›Let a larger Muse model plan and hand Ember bounded subtasks; switching between them only changes the model field

Quickstart

Python
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-ember-1.0",
    input="Classify this support ticket as billing, bug, or feature request: I was charged twice for my October invoice.",
)
print(response.output_text)

Meta's quickstart uses the OpenAI SDK pointed at the Meta Model API (pip install openai). The SDK does not read MODEL_API_KEY by itself, so pass it explicitly.

API model id: muse-ember-1.0

Benchmarks

BenchmarkScoreNotes
Terminal-Bench 4.040.9% (max effort)Meta-reported, mini-swe-agent harness; Claude Haiku 5.5 39.2% (Claude Code), GPT-6 Luna 16.4% (Codex)
Artificial Analysis Intelligence Index36Third-party (Intelligence Index v5, November 9, 2026), as published by Meta; Muse Spark 1.3 scores 37
AutomationBench-AA37.8% (max effort)Artificial Analysis run cited by Meta; GPT-6 Luna 20.2%, Claude Haiku 5.5 13.9%
FrontierCode 2.0-AA27.2% (max effort)Artificial Analysis run cited by Meta; Claude Haiku 5.5 34.9%, GPT-6 Luna 25.2%
OSWorld 2.169.5% (max effort)Meta internal evaluation, partial-credit score; Claude Haiku 5.5 68.9%, GPT-6 Luna 50.5%

Source: Meta - Introducing Muse Ember

Compare Muse Ember

Compare Muse Ember with any other model

Build a comparison →
All model comparisons →

Learn the concepts