822B
- Context
- 1Mtokens
- Released
- 3d ago
Cost/1M tokens
Muse Ember is a model from Meta that the launch announcement introduces as the company's smallest and most efficient model. Meta claims it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, matching Haiku's best score at about a third of the cost. It is built for complex reasoning execution and repetitive agent work such as tool calls, coding subtasks, routing, extraction, and long document parsing.
- Multimodal model
- Accepts images, audio, video, files
- 1M context
- Reasoning
- Tools
- Structured output
How it compares
Score vs. costArena WebDev
Each model at its best score · Dashed: the best score at each cost
- Meta
- Anthropic
- Google DeepMind
- OpenAI
- Moonshot AI
- xAI
- Xiaomi
- Z.ai
Vs. your stack
Add the models you use to see how this one compares.
Sign inAdditional source evidence
- Source model
- meta/muse-ember-1.0
- Source family
- muse
- Last updated
- 2026-11-09
- Model context
- 1M
- Model output limit
- 524K
- Open weights
- No
- File attachments
- Yes
- Reasoning
- Yes
- Tool calling
- Yes
- Temperature control
- Yes
Published benchmarks
| Benchmark | Score | Evaluation details | Evidence |
|---|---|---|---|
| AutomationBench | 37.8 · strict pass@1 | max effort; Artificial Analysis AutomationBench-AA · 2026-11-09 | Source |
| Agents' Last Exam | 61.6 · score | Meta internal evaluation framework · Linux subset · max effort · 2026-11-09 | Source |
| Terminal-Bench | 40.9 · accuracy | mini-swe-agent · max effort · 4.0 · 2026-11-09 | Source |
| FrontierCode | 27.2 · score | max effort; Artificial Analysis FrontierCode-AA · 2.0 · 2026-11-09 | Source |
| GDP.pdf | 22.8 · score | max effort; Artificial Analysis GDP.pdf-AA · 2026-11-09 | Source |
| OSWorld | 69.5 · mean partial score | max effort · 2.1 · 2026-11-09 | Source |
| OSWorld | 33.8 · binary completion rate | max effort · 2.1 · 2026-11-09 | Source |
| PerceptionBench | 80.6 · score | Meta internal evaluation framework · max effort; with tools · 2026-11-09 | Source |
| Artificial Analysis Intelligence Index | 36 · index score | max effort · 5 · 2026-11-09 | Source |
Official references
Complete source record (JSON)
{
"id": "meta/muse-ember-1.0",
"name": "Muse Ember",
"description": "Muse Ember is Meta's smallest and most efficient reasoning model, built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing. It pairs with Muse Wonder or Muse Spark, which plan while Muse Ember handles the volume.",
"family": "muse",
"last_updated": "2026-11-09",
"attachment": true,
"reasoning": true,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text",
"image",
"video",
"pdf",
"audio"
],
"output": [
"text"
]
},
"limit": {
"context": 1048576,
"output": 524288
},
"benchmarks": [
{
"name": "AutomationBench",
"score": 37.8,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "strict pass@1",
"variant": "max effort; Artificial Analysis AutomationBench-AA",
"date": "2026-11-09"
},
{
"name": "Agents' Last Exam",
"score": 61.6,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "score",
"harness": "Meta internal evaluation framework",
"variant": "max effort",
"dataset": "Linux subset",
"date": "2026-11-09"
},
{
"name": "Terminal-Bench",
"score": 40.9,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "accuracy",
"harness": "mini-swe-agent",
"variant": "max effort",
"version": "4.0",
"date": "2026-11-09"
},
{
"name": "FrontierCode",
"score": 27.2,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "score",
"variant": "max effort; Artificial Analysis FrontierCode-AA",
"version": "2.0",
"date": "2026-11-09"
},
{
"name": "GDP.pdf",
"score": 22.8,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "score",
"variant": "max effort; Artificial Analysis GDP.pdf-AA",
"date": "2026-11-09"
},
{
"name": "OSWorld",
"score": 69.5,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "mean partial score",
"variant": "max effort",
"version": "2.1",
"date": "2026-11-09"
},
{
"name": "OSWorld",
"score": 33.8,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "binary completion rate",
"variant": "max effort",
"version": "2.1",
"date": "2026-11-09"
},
{
"name": "PerceptionBench",
"score": 80.6,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "score",
"harness": "Meta internal evaluation framework",
"variant": "max effort; with tools",
"date": "2026-11-09"
},
{
"name": "Artificial Analysis Intelligence Index",
"score": 36,
"source": "https://research.meta.ai/blog/muse-ember",
"metric": "index score",
"variant": "max effort",
"version": "5",
"date": "2026-11-09"
}
],
"release_date": "2026-11-09",
"open_weights": false,
"links": [
{
"url": "https://research.meta.ai/blog/muse-ember",
"label": "Announcement",
"type": "announcement"
}
]
}Frequently asked questions
API and pricing
Input costs $0.10 per 1M tokens and output costs $0.50 per 1M tokens. Across 2 paid providers the input price range is $0.10 to $0.10 per 1M tokens.
Google is the cheapest provider for Muse Ember at the lowest blended price with a 3:1 input to output ratio. Google charges $0.10 per 1M input tokens and $0.50 per 1M output tokens.
Yes. Cached input costs $0.004 per 1M tokens versus $0.10 per 1M tokens uncached, a 96% discount, and cache writes cost $0.125 per 1M tokens.
A typical request of 10,000 input tokens and 1,000 output tokens costs $0.0015 at the cheapest provider, Google.
Of the 2 endpoints with uptime data in the last day, 1 ran at 99% uptime or better. The best uptime was 100.0%, by Meta.
Privacy and data
Google is the Muse Ember provider with zero data retention (ZDR eligible or 0-day retention).
Capabilities and specs
The context window is 1M tokens (1,048,576) and the maximum output is 944K tokens (943,718).
Yes. Muse Ember's input modalities are text, image, video, file and audio.
Yes, tool calling (function calling) is supported.
Yes, structured output (JSON schema or JSON mode) is supported.
Yes. Reasoning is supported and always on, with a default effort of max. Effort levels are max, xhigh, high, medium, low and minimal, exposed by 2 of 2 tracked providers, and a provider offers the reasoning_effort setting.
Quality and benchmarks
Muse Ember ranks 36 on Arena WebDev, Arena's own rank, with a score of 1609.
Matched on model kind, blended price and release date: Claude Haiku 5.5 by Anthropic at $0.20 per 1M tokens, GPT-6 Luna by OpenAI at $0.20 per 1M tokens, and Particle-1 by Particle AI at $0.1625 per 1M tokens.
Claude Haiku 5.5Muse Ember vs Claude Haiku 5.5GPT-6 LunaMuse Ember vs GPT-6 LunaParticle-1Muse Ember vs Particle-1
Kimi K3 is the best open-weight alternative, ranked 19 on Arena Text as the highest-ranked open-weight model. Muse Ember itself is closed.
Licensing and access
No. Muse Ember is closed.
Muse Ember is available on Google Vertex AI. There is no tracked endpoint on Azure or Amazon Bedrock.
Basics
Muse Ember was made by Meta and released on November 9, 2026.
No. Muse Ember is closed and the listed API price is $0.10 input and $0.50 output per 1M tokens.
Generated from Vertical Tabs data, updated November 11, 2026.
Discussion
No comments yet
Be the first to share your thoughts.