Muse 发布最小高效模型 Ember 主打智能体任务
Muse 推出其最小、最高效的模型 Muse Ember，面向工具调用、编码子任务、路由、信息抽取与长文档解析等智能体高频重复工作。官方称其在任意推理档位下单位成本智能均优于 Claude Haiku 5.5 与 GPT-6 Luna，并以约三分之一成本追平 Haiku 最佳成绩，处于 Artificial Analysis 榜单效率前沿。
报道正文
本文目录17 个章节
Today, we're introducing Muse Ember, our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.
Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting: it matches Haiku's best score at about a third of the cost and beats every Luna setting. It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.
Intelligence vs. Cost Per Task
Artificial Analysis Intelligence Index against cost per task at each reasoning effort, from the raw Artificial Analysis results. Muse Ember costs are estimates at proposed pricing.
- xhigh
- max
- medium
- high
- xhigh
- max
- high
- xhigh
- max
- Muse Ember
- Claude Haiku 5.5
- GPT-6 Luna
Intelligence vs. Cost Per Task: Intelligence vs. Cost Per Task
| Series | Measurement | Value |
|---|---|---|
| Muse Ember | Cost Per Task (USD) | $0.116 |
| Muse Ember | Score | 33.9 |
| Muse Ember | Cost Per Task (USD) | $0.136 |
| Muse Ember | Score | 35.8 |
| Claude Haiku 5.5 | Cost Per Task (USD) | $0.092 |
| Claude Haiku 5.5 | Score | 24.6 |
| Claude Haiku 5.5 | Cost Per Task (USD) | $0.159 |
| Claude Haiku 5.5 | Score | 29.6 |
| Claude Haiku 5.5 | Cost Per Task (USD) | $0.248 |
| Claude Haiku 5.5 | Score | 34.3 |
| Claude Haiku 5.5 | Cost Per Task (USD) | $0.411 |
| Claude Haiku 5.5 | Score | 36.2 |
| GPT-6 Luna | Cost Per Task (USD) | $0.073 |
| GPT-6 Luna | Score | 21.4 |
| GPT-6 Luna | Cost Per Task (USD) | $0.104 |
| GPT-6 Luna | Score | 23.3 |
| GPT-6 Luna | Cost Per Task (USD) | $0.152 |
| GPT-6 Luna | Score | 27.8 |
Artificial Analysis | November 9, 2026. Each point is one reasoning-effort setting. Muse Ember cost at list pricing.
Muse Ember evaluation results
Muse Ember evaluation results: Evaluation results| Evaluation results | Featured model.Muse Ember (max) | Claude Haiku 5.5 (max) | GPT-6 Luna (max) | | --- | --- | --- | --- | | Agents | AutomationBench-AA E2E business workflows | Best result.37.8 | 13.9 | 20.2 | | Agents' Last Exam End-to-end agent tasks | Best result.61.6 | 57.0 | 57.1 | | Personal Superintelligence Index Internal composite index of personal and knowledge work | Best result.46.8 | 36.1 | 37.0 | | Coding | Terminal-Bench 4.0 Agentic terminal coding, best published harness | Best result.40.9 | 39.2 | 16.4 | | Coding Agent Index-AA Coding in real agent harnesses | Best result.46 | 37 | 31 | | FrontierCode 2.0-AA Real engineering tasks judged on mergeability | 27.2 | Best result.34.9 | 25.2 | | Multimodal | GDP.pdf-AA Professional work over long documents | Best result.22.8 | 20.8 | Best result.22.8 | | OSWorld 2.1 Computer use across desktop apps | Best result.69.5(partial)33.8(binary) | 68.9(partial)27.8(binary) | 50.5(partial)17.0(binary) | | PerceptionBench (w/ tools)Fine-grained visual perception | Best result.80.6 | 78.6 | 62.8 | | General | Artificial Analysis Index | Best result.36 | Best result.36 | 28 |
For more details about our evaluations, see our report.
Muse Ember is available today on the Meta Model API, at $0.10 per million input tokens and $0.50 per million output tokens.
In early testing, our partners put Muse Ember to work on the kinds of tasks agents repeat all day, from reviewing contracts to writing code.
“On the Box AI Complex Work Eval, Muse Ember delivers impressive accuracy, speed and token efficiency. It can score a contract against a company's negotiation playbook clause by clause and land on the right overall risk rating, or screen candidates against a detailed brief and flag the borderline ones who still qualify as alternates. Muse Ember is a fast, efficient foundation for agents doing real work on enterprise content.”
Author Yashodha Bhavnani
Role VP of AI Products
Company Box
Agents
Agents spend most of their tokens on many small, structured steps, and Muse Ember is built for exactly that. It can carry a business workflow end to end, pulling records, calling tools in sequence, and updating systems across apps.
With a 1M-token context window and up to 512K output tokens, it can take in a full codebase or document set in a single call.
AutomationBench-AA
AutomationBench-AA results at max reasoning, reproduced from the doc's current chart.. Score
AutomationBench-AA: AutomationBench-AA
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 37.8% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 20.2% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 13.9% |
Artificial Analysis | November 9, 2026. Strict pass@1.
Agents' Last Exam
Agents' Last Exam results at max reasoning, reproduced from the doc's current chart.. Score
Agents' Last Exam: Agents' Last Exam
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 61.6% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 57.1% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 57.0% |
Meta internal evaluation. Linux subset.
Personal Superintelligence Index
Personal Superintelligence Index results at max reasoning, reproduced from the doc's current chart.. Score
Personal Superintelligence Index: Personal Superintelligence Index
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 46.8% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 37.0% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 36.1% |
Internal composite index of personal and knowledge work.
Coding
Muse Ember is built for the coding subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes.
Terminal-Bench 4.0
Terminal-Bench 4.0 results at max reasoning, reproduced from the doc's current chart.. Score
Terminal-Bench 4.0: Terminal-Bench 4.0
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 40.9% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 39.2% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 16.4% |
Each model in its best published harness. See footnote 1.
Coding Agent Index-AA
Coding Agent Index-AA results at max reasoning, reproduced from the doc's current chart.. Score
Coding Agent Index-AA: Coding Agent Index-AA
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 46 |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 37 |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 31 |
Artificial Analysis | November 9, 2026
FrontierCode 2.0-AA
FrontierCode 2.0-AA results at max reasoning, reproduced from the doc's current chart.. Score
FrontierCode 2.0-AA: FrontierCode 2.0-AA
| Series | Measurement | Value |
|---|---|---|
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 34.9% |
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 27.2% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 25.2% |
Artificial Analysis | November 9, 2026
For the hardest problems, pair it with Muse Wonder: Muse Wonder plans, Muse Ember handles the volume.
Multimodal
Muse Ember natively understands text, images, audio, and video, so agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. It can work through long real-world documents, operate a computer directly from screenshots, and pick out fine visual detail.
GDP.pdf-AA
GDP.pdf-AA results at max reasoning, reproduced from the doc's current chart.. Score
GDP.pdf-AA: GDP.pdf-AA
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 22.8% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 22.8% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 20.8% |
Artificial Analysis | November 9, 2026
OSWorld 2.1
OSWorld 2.1 results at max reasoning, reproduced from the doc's current chart.. Score
OSWorld 2.1: OSWorld 2.1
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 69.5% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 68.9% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 50.5% |
Meta internal evaluation. Partial-credit reward.
PerceptionBench
PerceptionBench results at max reasoning, reproduced from the doc's current chart.. Score
PerceptionBench: PerceptionBench
| Series | Measurement | Value |
|---|---|---|
| Muse Ember (max) | Model | Muse Ember (max) |
| Muse Ember (max) | Score | 80.6% |
| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |
| Claude Haiku 5.5 (max) | Score | 78.6% |
| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |
| GPT-6 Luna (max) | Score | 62.8% |
Meta internal evaluation. With tools.
Safety
Before release, we evaluated Muse Ember under our Meta Superintelligence Scaling Framework, including chemical and biological, cybersecurity, and loss-of-control risks. Across the categories assessed, our evaluations found that Muse Ember remained below the framework’s high-risk thresholds. Though it’s a small model, it received the same core safety training and testing. We also evaluated how Muse Ember behaves when acting on users' behalf. That covers whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents, or tool outputs that try to redirect it, and seeks human input before high-impact actions.
Availability
Muse Ember is available today on Muse Code and Meta Model API as muse-ember-1.0. It costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens. Muse Ember supports the Responses API.
When we introduced Muse Wonder and Muse Wonder Full, we said we would use them to develop specialized models we can bring to everyone. Muse Ember is the first trained using Muse Wonder Full. It rounds out our portfolio of models: Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale at about $0.14 per Artificial Analysis task, a cost low enough to run in every product, for everyone.
Muse Ember and Muse Spark 1.3 comparison
Muse Ember and Muse Spark 1.3 comparison: Artificial Analysis comparison| Artificial Analysis | Featured model.Muse Ember (max) | Featured model.Muse Spark 1.3 (max) | | --- | --- | --- | | Intelligence Index | 36 | Best result.37 | | Cost Per Intelligence Index Task | Best result.$0.14 | $2.39 | | Price Per 1M Input / Output Tokens | Best result.$0.10(input)$0.50(output) | $1.25(input)$4.25(output) |
Our goal is personal superintelligence: models that help you pursue your goals, create what you imagine, deepen your relationships, and take action on what you value most. Bringing it to everyone takes more than frontier capability. It takes frontier capability at a price that scales. We have even more efficient models in training and look forward to sharing what's next.