Meta AI released Muse Ember on Nov. 9, 2026^2,3. Nobody outside the lab has posted independent test scores for it yet⁴. The model is listed at $0.10 per million input tokens and $0.50 per million output tokens, with a 1,048,576-token context window^3,5. Meta Superintelligence Lab says it is the smallest and most efficient Muse model, built for repeated agent steps such as tool calls, coding subtasks, routing, extraction and long-document parsing^2,3.

The field it joins is already ranked. Anthropic's Claude Fable 5.2, also dated Nov. 9, sits first of 44 on Epoch AI's Capabilities Index at 174.2⁴. OpenAI's GPT-6 Astra is second at 169.2; Google DeepMind's Gemini 4 Argon is seventh at 166.5⁴. Meta's previous Muse Wonder, released Oct. 12, ranks 15th at 162 on that index⁴. Ember has no Epoch rank⁴.

Meta's own tables, not Epoch AI's, compare Ember at max reasoning with Claude Haiku 5.5 and GPT-6 Luna³. Meta reports 37.8% on AutomationBench-AA versus 13.9% for Haiku 5.5 and 20.2% for Luna; 61.6% on Agents' Last Exam versus 57.0% and 57.1%; and 40.9% on Terminal-Bench 4.0 versus 39.2% and 16.4%³. On FrontierCode 2.0-AA, Meta reports Ember at 27.2%, behind Haiku 5.5 at 34.9% and ahead of Luna at 25.2%³.

Meta says Ember matches Haiku's best Artificial Analysis score at about a third of the cost and sits on that vendor's efficiency frontier³. TechCrunch reported the same cost-per-task framing and the line that Muse Wonder plans while Muse Ember handles the volume². Those comparisons are Meta's, against mid-tier models, not against Fable 5.2, Astra or Argon^2,3.

What is not known is Ember's score on GPQA Diamond, FrontierMath, ARC-AGI-2, SimpleQA Verified, SciCode, APEX-Agents or CritPt as run by Epoch AI, the ARC Prize Foundation, Artificial Analysis, Mercor or CritPt on the ranked rivals⁴.