Meta Superintelligence Labs released Muse Ember, a hosted multimodal reasoning model with a 1-million-token context window and 512K-token output capacity, designed for high-volume agent subtasks such as tool calls, coding, routing, extraction, and long-document parsing. Trained using Meta's largest model, Muse Wonder Full, Ember natively processes text, images, audio, and video, enabling agents to interpret screenshots, charts, scanned pages, voice memos, and meeting recordings while operating desktop apps from screenshots. On the Artificial Analysis Intelligence Index at max reasoning effort, Ember scores 36, tying Claude Haiku 5.5 and leading GPT-6 Luna's 28. It excels on AutomationBench-AA with 37.8% task completion across 657 Zapier workflows versus Haiku's 13.9% and Luna's 20.2%, leads Terminal-Bench 4.0 at 40.9 versus Haiku's 39.2, and scores 80.6 on PerceptionBench with tools against Haiku's 78.6. However, it trails on FrontierCode 2.0-AA at 27.2 versus Haiku's 34.9. Despite identical base token pricing ($0.10/$0.50 per million input/output), Ember achieves roughly one-third the cost per task ($0.136 vs $0.411) due to token efficiency, with no long-context surcharges unlike rivals. Weights remain closed; access is via Meta Model API and Muse Code, with early partners including Box, Cline, Oracle, Cloudflare, and You.com.