Muse Ember

Muse Ember is Meta Superintelligence Labs' smallest and most efficient Muse model, released on 9 November 2026. It is built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long-document parsing.

Meta says it 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 has a 1M-token context window and up to 512K output tokens. List pricing is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. The model id is muse-ember-1.0 on Muse Code and the Meta Model API, and it supports the Responses API. OpenRouter serves it as meta/muse-ember and lists text, image, video, audio, and PDF inputs.

Highlights:

• Announced 2026-11-09; official post https://research.meta.ai/blog/muse-ember ; OpenRouter page https://openrouter.ai/meta/muse-ember ; covered the same day by TechCrunch;
• Positioned as the cheapest/fastest execution model in the Muse lineup, meant for subagents and high-throughput steps;
• Meta's line is “Muse Wonder plans, Muse Ember handles the volume”;
• First specialized model trained using Muse Wonder Full; Meta describes the portfolio as Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks at scale;
• Official comparison puts cost per Artificial Analysis task at about $0.14, versus about $2.39 for Muse Spark 1.3.

Why it matters

Agents spend most of their tokens on many small, structured steps. Meta put Ember in that gap: let a stronger model such as Muse Wonder plan, then send repeatable work—contract review, document analysis, multi-repo investigation—to a cheaper execution model. Official charts claim 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, and sitting on the efficiency frontier of the Artificial Analysis leaderboard. In early partner notes, Cloudflare said Ember ran up to 30% faster at the tail than the comparable model they tested; You.com reported a 2.2-second p95 on coding quests and 36–48% fewer output tokens.

Selected public numbers (max reasoning, from Meta’s post and Artificial Analysis on 9 November 2026): AutomationBench-AA 37.8 (Haiku 13.9, Luna 20.2); Agents’ Last Exam 61.6; Terminal-Bench 4.0 40.9; Coding Agent Index-AA 46. The important caveat: on FrontierCode 2.0-AA, which scores whether code changes would be merged, Haiku 5.5 led at 34.9% versus Ember’s 27.2%. Ember is not ahead on every coding eval.

Limits

Ember is not the flagship reasoning model. Meta itself says to pair it with Muse Wonder on the hardest problems. Heavier merge-style engineering tasks may still favor a larger model. Price, context, and multimodal limits follow the current Meta Model API / Muse Code / OpenRouter listings; check docs for region, quota, and accepted input types. On safety, Meta says evaluations under its Superintelligence Scaling Framework stayed below high-risk thresholds for chemical and biological, cybersecurity, and loss-of-control risks—that is the lab’s own assessment, not a third-party certification. Do not treat unpublished parameter counts or private benchmark scores as facts.

1. Open https://research.meta.ai/blog/muse-ember or https://openrouter.ai/meta/muse-ember and confirm the current model id (muse-ember-1.0 on Meta; meta/muse-ember on OpenRouter) and list price;
2. Call it from Muse Code, the Meta Model API (Responses API is supported), or OpenRouter’s compatible endpoint;
3. Use it as a subagent in a harness: leave planning to a stronger model and send tool calls, extraction, targeted fixes, and long-document parsing to Ember;
4. Before production use, check license, regional availability, cached-input pricing, and eval methodology; do not assume unpublished quotas or parameter counts.