{"article":{"slug":"introducing-our-smallest-most-efficient-model-muse-ember","title":"Introducing Our Smallest, Most Efficient Model: Muse Ember","subtitle":null,"summary":"Meta introduces Muse Ember, its smallest and most efficient model, built for agent workloads like tool calls, coding subtasks, routing, extraction and long document parsing. Meta says it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna.","content_type":"announcement","language":"en","canonical_url":"https://research.meta.ai/blog/muse-ember","author":{"name":"Meta AI Research","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"Meta AI Research","url":"https://research.meta.ai/","listing_slug":null,"listing":null},"topics":[{"name":"Artificial Intelligence","slug":"artificial-intelligence","url":"https://listedarticles.com/topics/artificial-intelligence"},{"name":"LLMs","slug":"llms","url":"https://listedarticles.com/topics/llms"},{"name":"Meta","slug":"meta","url":"https://listedarticles.com/topics/meta"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":1996,"reading_minutes":9,"published_at":"2026-11-09T00:00:00.000Z","added_at":"2026-11-09T20:07:11.516Z","updated_at":"2026-11-09T20:07:11.516Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/introducing-our-smallest-most-efficient-model-muse-ember","markdown_url":"https://listedarticles.com/articles/introducing-our-smallest-most-efficient-model-muse-ember.md","example":false,"citation":"Meta AI Research, Meta AI Research. \"Introducing Our Smallest, Most Efficient Model: Muse Ember.\" 9 Nov 2026. https://research.meta.ai/blog/muse-ember (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://research.meta.ai/blog/muse-ember"},"body_markdown":"# Introducing Our Smallest, Most Efficient Model: Muse Ember\n\nNovember 9, 2026·3 minute read\n\nToday, 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.\n\nMuse 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.\n\n### Intelligence vs. Cost Per Task\n\nArtificial 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.\n\n- xhigh\n- max\n- medium\n- high\n- xhigh\n- max\n- high\n- xhigh\n- max\n- Muse Ember\n- Claude Haiku 5.5\n- GPT-6 Luna\n\nIntelligence vs. Cost Per Task: Intelligence vs. Cost Per Task\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember | Cost Per Task (USD) | $0.116 |\n| Muse Ember | Score | 33.9 |\n| Muse Ember | Cost Per Task (USD) | $0.136 |\n| Muse Ember | Score | 35.8 |\n| Claude Haiku 5.5 | Cost Per Task (USD) | $0.092 |\n| Claude Haiku 5.5 | Score | 24.6 |\n| Claude Haiku 5.5 | Cost Per Task (USD) | $0.159 |\n| Claude Haiku 5.5 | Score | 29.6 |\n| Claude Haiku 5.5 | Cost Per Task (USD) | $0.248 |\n| Claude Haiku 5.5 | Score | 34.3 |\n| Claude Haiku 5.5 | Cost Per Task (USD) | $0.411 |\n| Claude Haiku 5.5 | Score | 36.2 |\n| GPT-6 Luna | Cost Per Task (USD) | $0.073 |\n| GPT-6 Luna | Score | 21.4 |\n| GPT-6 Luna | Cost Per Task (USD) | $0.104 |\n| GPT-6 Luna | Score | 23.3 |\n| GPT-6 Luna | Cost Per Task (USD) | $0.152 |\n| GPT-6 Luna | Score | 27.8 |\n\nArtificial Analysis | November 9, 2026. Each point is one reasoning-effort setting. Muse Ember cost at list pricing.\n\n### Muse Ember evaluation results\n\nMuse Ember evaluation results: Evaluation results\n\n| Evaluation results | | Featured model. Muse Ember (max) | Claude Haiku 5.5 (max) | GPT-6 Luna (max) |\n| --- | --- | --- | --- | --- |\n| Agents | AutomationBench-AAE2E business workflows | Best result. 37.8 | 13.9 | 20.2 |\n| Agents' Last ExamEnd-to-end agent tasks | Best result. 61.6 | 57.0 | 57.1 |\n| Personal Superintelligence IndexInternal composite index of personal and knowledge work | Best result. 46.8 | 36.1 | 37.0 |\n| Coding | Terminal-Bench 4.0Agentic terminal coding, best published harness | Best result. 40.9 | 39.2 | 16.4 |\n| Coding Agent Index-AACoding in real agent harnesses | Best result. 46 | 37 | 31 |\n| FrontierCode 2.0-AAReal engineering tasks judged on mergeability | 27.2 | Best result. 34.9 | 25.2 |\n| Multimodal | GDP.pdf-AAProfessional work over long documents | Best result. 22.8 | 20.8 | Best result. 22.8 |\n| OSWorld 2.1Computer use across desktop apps | Best result. 69.5(partial)33.8(binary) | 68.9(partial)27.8(binary) | 50.5(partial)17.0(binary) |\n| PerceptionBench (w/ tools)Fine-grained visual perception | Best result. 80.6 | 78.6 | 62.8 |\n| General | Artificial Analysis Index | Best result. 36 | Best result. 36 | 28 |\n\nFor more details about our evaluations, see our [report](https://research.meta.ai/static/muse-ember-evaluation-report \"Muse Ember Evaluation Report\").\n\nMuse Ember is available today on the Meta Model API, at $0.10 per million input tokens and $0.50 per million output tokens.\n\nIn early testing, our partners put Muse Ember to work on the kinds of tasks agents repeat all day, from reviewing contracts to writing code.\n\n![Box logo](https://research.meta.ai/articles/muse-glint/partner-quotes/box-logo-light-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1)\n\n> “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.”\n\nAuthor\n:   Yashodha Bhavnani\n\nRole\n:   VP of AI Products\n\nCompany\n:   Box\n\n![Cline logo](https://research.meta.ai/articles/muse-glint/partner-quotes/cline-logo-light-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1)\n\n> “It's incredible seeing Meta release a lightweight model for developers who want a fast, low-cost model they can rely on, and for teams cutting costs by running it as the subagent in their harness. Muse Ember is purpose-built for, and excels at, subagent work like coding and document analysis.”\n\nAuthor\n:   Saoud Rizwan\n\nRole\n:   Founder\n\nCompany\n:   Cline\n\n![Oracle logo](https://research.meta.ai/articles/muse-glint/partner-quotes/oracle-logo-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1)\n\n> “Muse Ember showed an encouraging ability to tackle complex engineering tasks in our early testing. We see strong potential for it to become a valuable tool for engineering teams.”\n\nAuthor\n:   Konstantinos Papamiltiadis (kp)\n\nRole\n:   SVP, Global Partnerships\n\nCompany\n:   Oracle\n\n![Cloudflare logo](https://research.meta.ai/articles/muse-glint/partner-quotes/cloudflare-logo-light-v2.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1)\n\n> “What stood out in our testing was speed. Muse Ember ran up to 30% faster at the tail than the comparable model we tested, and it handled broad, multi-repo investigation tasks well. It’s a strong fit for fast, high-throughput agentic workloads.”\n\nAuthor\n:   Dane Knecht\n\nRole\n:   CTO\n\nCompany\n:   Cloudflare\n\n![You.com logo](https://research.meta.ai/articles/muse-glint/partner-quotes/you-com-logo-light-v2.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1)\n\n> “Muse Ember completed every coding quest in our testing with the lowest latency of any model we tested: a p95 of 2.2 seconds, about 4x to 13x faster than leading open-weight models, and with 36–48% fewer output tokens. We’re excited about what Meta Superintelligence Labs is building next and about bringing it to our customers.”\n\nAuthor\n:   Chak Pothina\n\nRole\n:   Inference Lead\n\nCompany\n:   You.com\n\n## Agents\n\nAgents 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.\n\nWith 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.\n\n### AutomationBench-AA\n\nAutomationBench-AA results at max reasoning, reproduced from the doc's current chart.. Score\n\nAutomationBench-AA: AutomationBench-AA\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 37.8% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 20.2% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 13.9% |\n\nArtificial Analysis | November 9, 2026. Strict pass@1.\n\n### Agents' Last Exam\n\nAgents' Last Exam results at max reasoning, reproduced from the doc's current chart.. Score\n\nAgents' Last Exam: Agents' Last Exam\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 61.6% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 57.1% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 57.0% |\n\nMeta internal evaluation. Linux subset.\n\n### Personal Superintelligence Index\n\nPersonal Superintelligence Index results at max reasoning, reproduced from the doc's current chart.. Score\n\nPersonal Superintelligence Index: Personal Superintelligence Index\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 46.8% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 37.0% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 36.1% |\n\nInternal composite index of personal and knowledge work.\n\n## Coding\n\nMuse Ember is built for the coding subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes.\n\n### Terminal-Bench 4.0\n\nTerminal-Bench 4.0 results at max reasoning, reproduced from the doc's current chart.. Score\n\nTerminal-Bench 4.0: Terminal-Bench 4.0\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 40.9% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 39.2% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 16.4% |\n\nEach model in its best published harness. See footnote 1.\n\n### Coding Agent Index-AA\n\nCoding Agent Index-AA results at max reasoning, reproduced from the doc's current chart.. Score\n\nCoding Agent Index-AA: Coding Agent Index-AA\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 46 |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 37 |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 31 |\n\nArtificial Analysis | November 9, 2026\n\n### FrontierCode 2.0-AA\n\nFrontierCode 2.0-AA results at max reasoning, reproduced from the doc's current chart.. Score\n\nFrontierCode 2.0-AA: FrontierCode 2.0-AA\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 34.9% |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 27.2% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 25.2% |\n\nArtificial Analysis | November 9, 2026\n\nFor the hardest problems, pair it with Muse Wonder: Muse Wonder plans, Muse Ember handles the volume.\n\n## Multimodal\n\nMuse 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.\n\n### GDP.pdf-AA\n\nGDP.pdf-AA results at max reasoning, reproduced from the doc's current chart.. Score\n\nGDP.pdf-AA: GDP.pdf-AA\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 22.8% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 22.8% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 20.8% |\n\nArtificial Analysis | November 9, 2026\n\n### OSWorld 2.1\n\nOSWorld 2.1 results at max reasoning, reproduced from the doc's current chart.. Score\n\nOSWorld 2.1: OSWorld 2.1\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 69.5% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 68.9% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 50.5% |\n\nMeta internal evaluation. Partial-credit reward.\n\n### PerceptionBench\n\nPerceptionBench results at max reasoning, reproduced from the doc's current chart.. Score\n\nPerceptionBench: PerceptionBench\n\n| Series | Measurement | Value |\n| --- | --- | --- |\n| Muse Ember (max) | Model | Muse Ember (max) |\n| Muse Ember (max) | Score | 80.6% |\n| Claude Haiku 5.5 (max) | Model | Claude Haiku 5.5 (max) |\n| Claude Haiku 5.5 (max) | Score | 78.6% |\n| GPT-6 Luna (max) | Model | GPT-6 Luna (max) |\n| GPT-6 Luna (max) | Score | 62.8% |\n\nMeta internal evaluation. With tools.\n\n## Safety\n\nBefore 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.\n\n## Availability\n\nMuse 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.\n\nWhen 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.\n\n### Muse Ember and Muse Spark 1.3 comparison\n\nMuse Ember and Muse Spark 1.3 comparison: Artificial Analysis comparison\n\n| Artificial Analysis | Featured model. Muse Ember (max) | Featured model. Muse Spark 1.3 (max) |\n| --- | --- | --- |\n| Intelligence Index | 36 | Best result. 37 |\n| Cost Per Intelligence Index Task | Best result. $0.14 | $2.39 |\n| Price Per 1M Input / Output Tokens | Best result. $0.10(input)$0.50(output) | $1.25(input)$4.25(output) |\n\nOur 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.\n\n[Try Muse Ember today](https://dev.meta.ai)\n1. All models run at max reasoning effort. Artificial Analysis results are from Intelligence Index v5, as published November 9, 2026; cost per task uses each model's list pricing. Internal results are from Meta's evaluation suite. The Personal Superintelligence Index (PSI) is an internal composite of eight evaluations: Agents' Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx and VBT. Terminal-Bench 4.0 uses each model's best published harness: mini-swe-agent (Muse Ember), Claude Code (Claude Haiku 5.5), and Codex (GPT-6 Luna). OSWorld 2.1 uses partial-credit reward; PerceptionBench is scored with tools.\n","body_html":"<h1 id=\"introducing-our-smallest-most-efficient-model-muse-ember\">Introducing Our Smallest, Most Efficient Model: Muse Ember</h1>\n<p>November 9, 2026·3 minute read</p>\n<p>Today, we&#39;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.</p>\n<p>Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting: it matches Haiku&#39;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.</p>\n<h3 id=\"intelligence-vs-cost-per-task\">Intelligence vs. Cost Per Task</h3>\n<p>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.</p>\n<ul><li>xhigh</li><li>max</li><li>medium</li><li>high</li><li>xhigh</li><li>max</li><li>high</li><li>xhigh</li><li>max</li><li>Muse Ember</li><li>Claude Haiku 5.5</li><li>GPT-6 Luna</li></ul>\n<p>Intelligence vs. Cost Per Task: Intelligence vs. Cost Per Task</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember</td><td>Cost Per Task (USD)</td><td>$0.116</td></tr><tr><td>Muse Ember</td><td>Score</td><td>33.9</td></tr><tr><td>Muse Ember</td><td>Cost Per Task (USD)</td><td>$0.136</td></tr><tr><td>Muse Ember</td><td>Score</td><td>35.8</td></tr><tr><td>Claude Haiku 5.5</td><td>Cost Per Task (USD)</td><td>$0.092</td></tr><tr><td>Claude Haiku 5.5</td><td>Score</td><td>24.6</td></tr><tr><td>Claude Haiku 5.5</td><td>Cost Per Task (USD)</td><td>$0.159</td></tr><tr><td>Claude Haiku 5.5</td><td>Score</td><td>29.6</td></tr><tr><td>Claude Haiku 5.5</td><td>Cost Per Task (USD)</td><td>$0.248</td></tr><tr><td>Claude Haiku 5.5</td><td>Score</td><td>34.3</td></tr><tr><td>Claude Haiku 5.5</td><td>Cost Per Task (USD)</td><td>$0.411</td></tr><tr><td>Claude Haiku 5.5</td><td>Score</td><td>36.2</td></tr><tr><td>GPT-6 Luna</td><td>Cost Per Task (USD)</td><td>$0.073</td></tr><tr><td>GPT-6 Luna</td><td>Score</td><td>21.4</td></tr><tr><td>GPT-6 Luna</td><td>Cost Per Task (USD)</td><td>$0.104</td></tr><tr><td>GPT-6 Luna</td><td>Score</td><td>23.3</td></tr><tr><td>GPT-6 Luna</td><td>Cost Per Task (USD)</td><td>$0.152</td></tr><tr><td>GPT-6 Luna</td><td>Score</td><td>27.8</td></tr></tbody></table></div>\n<p>Artificial Analysis | November 9, 2026. Each point is one reasoning-effort setting. Muse Ember cost at list pricing.</p>\n<h3 id=\"muse-ember-evaluation-results\">Muse Ember evaluation results</h3>\n<p>Muse Ember evaluation results: Evaluation results</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Evaluation results</th><th></th><th>Featured model. Muse Ember (max)</th><th>Claude Haiku 5.5 (max)</th><th>GPT-6 Luna (max)</th></tr></thead><tbody><tr><td>Agents</td><td>AutomationBench-AAE2E business workflows</td><td>Best result. 37.8</td><td>13.9</td><td>20.2</td></tr><tr><td>Agents&#39; Last ExamEnd-to-end agent tasks</td><td>Best result. 61.6</td><td>57.0</td><td>57.1</td><td></td></tr><tr><td>Personal Superintelligence IndexInternal composite index of personal and knowledge work</td><td>Best result. 46.8</td><td>36.1</td><td>37.0</td><td></td></tr><tr><td>Coding</td><td>Terminal-Bench 4.0Agentic terminal coding, best published harness</td><td>Best result. 40.9</td><td>39.2</td><td>16.4</td></tr><tr><td>Coding Agent Index-AACoding in real agent harnesses</td><td>Best result. 46</td><td>37</td><td>31</td><td></td></tr><tr><td>FrontierCode 2.0-AAReal engineering tasks judged on mergeability</td><td>27.2</td><td>Best result. 34.9</td><td>25.2</td><td></td></tr><tr><td>Multimodal</td><td>GDP.pdf-AAProfessional work over long documents</td><td>Best result. 22.8</td><td>20.8</td><td>Best result. 22.8</td></tr><tr><td>OSWorld 2.1Computer use across desktop apps</td><td>Best result. 69.5(partial)33.8(binary)</td><td>68.9(partial)27.8(binary)</td><td>50.5(partial)17.0(binary)</td><td></td></tr><tr><td>PerceptionBench (w/ tools)Fine-grained visual perception</td><td>Best result. 80.6</td><td>78.6</td><td>62.8</td><td></td></tr><tr><td>General</td><td>Artificial Analysis Index</td><td>Best result. 36</td><td>Best result. 36</td><td>28</td></tr></tbody></table></div>\n<p>For more details about our evaluations, see our <a href=\"https://research.meta.ai/static/muse-ember-evaluation-report\" title=\"Muse Ember Evaluation Report\" rel=\"nofollow ugc noopener\">report</a>.</p>\n<p>Muse Ember is available today on the Meta Model API, at $0.10 per million input tokens and $0.50 per million output tokens.</p>\n<p>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.</p>\n<figure><img src=\"https://research.meta.ai/articles/muse-glint/partner-quotes/box-logo-light-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1\" alt=\"Box logo\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<blockquote><p>“On the Box AI Complex Work Eval, Muse Ember delivers impressive accuracy, speed and token efficiency. It can score a contract against a company&#39;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.”</p></blockquote>\n<p>Author\n:   Yashodha Bhavnani</p>\n<p>Role\n:   VP of AI Products</p>\n<p>Company\n:   Box</p>\n<figure><img src=\"https://research.meta.ai/articles/muse-glint/partner-quotes/cline-logo-light-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1\" alt=\"Cline logo\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<blockquote><p>“It&#39;s incredible seeing Meta release a lightweight model for developers who want a fast, low-cost model they can rely on, and for teams cutting costs by running it as the subagent in their harness. Muse Ember is purpose-built for, and excels at, subagent work like coding and document analysis.”</p></blockquote>\n<p>Author\n:   Saoud Rizwan</p>\n<p>Role\n:   Founder</p>\n<p>Company\n:   Cline</p>\n<figure><img src=\"https://research.meta.ai/articles/muse-glint/partner-quotes/oracle-logo-v1.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1\" alt=\"Oracle logo\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<blockquote><p>“Muse Ember showed an encouraging ability to tackle complex engineering tasks in our early testing. We see strong potential for it to become a valuable tool for engineering teams.”</p></blockquote>\n<p>Author\n:   Konstantinos Papamiltiadis (kp)</p>\n<p>Role\n:   SVP, Global Partnerships</p>\n<p>Company\n:   Oracle</p>\n<figure><img src=\"https://research.meta.ai/articles/muse-glint/partner-quotes/cloudflare-logo-light-v2.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1\" alt=\"Cloudflare logo\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<blockquote><p>“What stood out in our testing was speed. Muse Ember ran up to 30% faster at the tail than the comparable model we tested, and it handled broad, multi-repo investigation tasks well. It’s a strong fit for fast, high-throughput agentic workloads.”</p></blockquote>\n<p>Author\n:   Dane Knecht</p>\n<p>Role\n:   CTO</p>\n<p>Company\n:   Cloudflare</p>\n<figure><img src=\"https://research.meta.ai/articles/muse-glint/partner-quotes/you-com-logo-light-v2.svg?dpl=dpl_ANtafPPBTB3QMKUEjE5giciZF6j1\" alt=\"You.com logo\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<blockquote><p>“Muse Ember completed every coding quest in our testing with the lowest latency of any model we tested: a p95 of 2.2 seconds, about 4x to 13x faster than leading open-weight models, and with 36–48% fewer output tokens. We’re excited about what Meta Superintelligence Labs is building next and about bringing it to our customers.”</p></blockquote>\n<p>Author\n:   Chak Pothina</p>\n<p>Role\n:   Inference Lead</p>\n<p>Company\n:   You.com</p>\n<h2 id=\"agents\">Agents</h2>\n<p>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.</p>\n<p>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.</p>\n<h3 id=\"automationbench-aa\">AutomationBench-AA</h3>\n<p>AutomationBench-AA results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>AutomationBench-AA: AutomationBench-AA</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>37.8%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>20.2%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>13.9%</td></tr></tbody></table></div>\n<p>Artificial Analysis | November 9, 2026. Strict pass@1.</p>\n<h3 id=\"agents-last-exam\">Agents&#39; Last Exam</h3>\n<p>Agents&#39; Last Exam results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>Agents&#39; Last Exam: Agents&#39; Last Exam</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>61.6%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>57.1%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>57.0%</td></tr></tbody></table></div>\n<p>Meta internal evaluation. Linux subset.</p>\n<h3 id=\"personal-superintelligence-index\">Personal Superintelligence Index</h3>\n<p>Personal Superintelligence Index results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>Personal Superintelligence Index: Personal Superintelligence Index</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>46.8%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>37.0%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>36.1%</td></tr></tbody></table></div>\n<p>Internal composite index of personal and knowledge work.</p>\n<h2 id=\"coding\">Coding</h2>\n<p>Muse Ember is built for the coding subtasks an agent sends off hundreds of times a session: reading files, running tests, making targeted fixes.</p>\n<h3 id=\"terminal-bench-4-0\">Terminal-Bench 4.0</h3>\n<p>Terminal-Bench 4.0 results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>Terminal-Bench 4.0: Terminal-Bench 4.0</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>40.9%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>39.2%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>16.4%</td></tr></tbody></table></div>\n<p>Each model in its best published harness. See footnote 1.</p>\n<h3 id=\"coding-agent-index-aa\">Coding Agent Index-AA</h3>\n<p>Coding Agent Index-AA results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>Coding Agent Index-AA: Coding Agent Index-AA</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>46</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>37</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>31</td></tr></tbody></table></div>\n<p>Artificial Analysis | November 9, 2026</p>\n<h3 id=\"frontiercode-2-0-aa\">FrontierCode 2.0-AA</h3>\n<p>FrontierCode 2.0-AA results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>FrontierCode 2.0-AA: FrontierCode 2.0-AA</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>34.9%</td></tr><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>27.2%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>25.2%</td></tr></tbody></table></div>\n<p>Artificial Analysis | November 9, 2026</p>\n<p>For the hardest problems, pair it with Muse Wonder: Muse Wonder plans, Muse Ember handles the volume.</p>\n<h2 id=\"multimodal\">Multimodal</h2>\n<p>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.</p>\n<h3 id=\"gdp-pdf-aa\">GDP.pdf-AA</h3>\n<p>GDP.pdf-AA results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>GDP.pdf-AA: GDP.pdf-AA</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>22.8%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>22.8%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>20.8%</td></tr></tbody></table></div>\n<p>Artificial Analysis | November 9, 2026</p>\n<h3 id=\"osworld-2-1\">OSWorld 2.1</h3>\n<p>OSWorld 2.1 results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>OSWorld 2.1: OSWorld 2.1</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>69.5%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>68.9%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>50.5%</td></tr></tbody></table></div>\n<p>Meta internal evaluation. Partial-credit reward.</p>\n<h3 id=\"perceptionbench\">PerceptionBench</h3>\n<p>PerceptionBench results at max reasoning, reproduced from the doc&#39;s current chart.. Score</p>\n<p>PerceptionBench: PerceptionBench</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Series</th><th>Measurement</th><th>Value</th></tr></thead><tbody><tr><td>Muse Ember (max)</td><td>Model</td><td>Muse Ember (max)</td></tr><tr><td>Muse Ember (max)</td><td>Score</td><td>80.6%</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Model</td><td>Claude Haiku 5.5 (max)</td></tr><tr><td>Claude Haiku 5.5 (max)</td><td>Score</td><td>78.6%</td></tr><tr><td>GPT-6 Luna (max)</td><td>Model</td><td>GPT-6 Luna (max)</td></tr><tr><td>GPT-6 Luna (max)</td><td>Score</td><td>62.8%</td></tr></tbody></table></div>\n<p>Meta internal evaluation. With tools.</p>\n<h2 id=\"safety\">Safety</h2>\n<p>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&#39; 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.</p>\n<h2 id=\"availability\">Availability</h2>\n<p>Muse Ember is available today on Muse Code and Meta Model API as <code>muse-ember-1.0</code>. 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.</p>\n<p>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.</p>\n<h3 id=\"muse-ember-and-muse-spark-1-3-comparison\">Muse Ember and Muse Spark 1.3 comparison</h3>\n<p>Muse Ember and Muse Spark 1.3 comparison: Artificial Analysis comparison</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Artificial Analysis</th><th>Featured model. Muse Ember (max)</th><th>Featured model. Muse Spark 1.3 (max)</th></tr></thead><tbody><tr><td>Intelligence Index</td><td>36</td><td>Best result. 37</td></tr><tr><td>Cost Per Intelligence Index Task</td><td>Best result. $0.14</td><td>$2.39</td></tr><tr><td>Price Per 1M Input / Output Tokens</td><td>Best result. $0.10(input)$0.50(output)</td><td>$1.25(input)$4.25(output)</td></tr></tbody></table></div>\n<p>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&#39;s next.</p>\n<p><a href=\"https://dev.meta.ai\" rel=\"nofollow ugc noopener\">Try Muse Ember today</a></p>\n<ol><li>All models run at max reasoning effort. Artificial Analysis results are from Intelligence Index v5, as published November 9, 2026; cost per task uses each model&#39;s list pricing. Internal results are from Meta&#39;s evaluation suite. The Personal Superintelligence Index (PSI) is an internal composite of eight evaluations: Agents&#39; Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx and VBT. Terminal-Bench 4.0 uses each model&#39;s best published harness: mini-swe-agent (Muse Ember), Claude Code (Claude Haiku 5.5), and Codex (GPT-6 Luna). OSWorld 2.1 uses partial-credit reward; PerceptionBench is scored with tools.</li></ol>","headings":[{"level":1,"text":"Introducing Our Smallest, Most Efficient Model: Muse Ember","id":"introducing-our-smallest-most-efficient-model-muse-ember"},{"level":3,"text":"Intelligence vs. Cost Per Task","id":"intelligence-vs-cost-per-task"},{"level":3,"text":"Muse Ember evaluation results","id":"muse-ember-evaluation-results"},{"level":2,"text":"Agents","id":"agents"},{"level":3,"text":"AutomationBench-AA","id":"automationbench-aa"},{"level":3,"text":"Agents' Last Exam","id":"agents-last-exam"},{"level":3,"text":"Personal Superintelligence Index","id":"personal-superintelligence-index"},{"level":2,"text":"Coding","id":"coding"},{"level":3,"text":"Terminal-Bench 4.0","id":"terminal-bench-4-0"},{"level":3,"text":"Coding Agent Index-AA","id":"coding-agent-index-aa"},{"level":3,"text":"FrontierCode 2.0-AA","id":"frontiercode-2-0-aa"},{"level":2,"text":"Multimodal","id":"multimodal"},{"level":3,"text":"GDP.pdf-AA","id":"gdp-pdf-aa"},{"level":3,"text":"OSWorld 2.1","id":"osworld-2-1"},{"level":3,"text":"PerceptionBench","id":"perceptionbench"},{"level":2,"text":"Safety","id":"safety"},{"level":2,"text":"Availability","id":"availability"},{"level":3,"text":"Muse Ember and Muse Spark 1.3 comparison","id":"muse-ember-and-muse-spark-1-3-comparison"}]}}