Meta introduced Muse Ember on 2026-11-09, calling it its “smallest and most efficient model” and aiming it at the repeated work inside agent systems: tool calls, coding subtasks, routing, extraction and long-document parsing. It is available on Muse Code and the Meta Model API as muse-ember-1.0, and Meta says Muse Ember, Muse Spark and Muse Wonder are now available on Google Cloud and Oracle Cloud. It natively handles text, images, audio and video, with a 1M-token context window and up to 512K output tokens.
Pricing for prompts up to 200K tokens is 0.10 dollars per million input tokens, 0.004 dollars per million cached input tokens and 0.50 dollars per million output tokens, tripling for longer prompts. Meta lists its cost per Artificial Analysis task at 0.14 dollars. Ember is the first model trained using Muse Wonder Full, and Meta frames the pairing directly: “Muse Wonder plans, Muse Ember handles the volume.”
Meta’s comparison table, at max reasoning, sets Ember against GPT-6 Luna and Claude Haiku 5.5. Ember reports 37.8 percent on AutomationBench-AA (Luna 20.2, Haiku 13.9), 61.6 percent on Agents’ Last Exam (Luna 57.1, Haiku 57.0), 40.9 percent on Terminal-Bench 4.0 (Haiku 39.2) and 69.5 percent on OSWorld 2.1. It trails Haiku 5.5 on FrontierCode 2.0-AA, 27.2 against 34.9 percent, and scores 36 on the Artificial Analysis Index against 37 for Muse Spark 1.3.
Why it matters: the small-model tier is now being sold explicitly as the worker half of a planner-worker agent stack, and Meta is pricing it well under its rivals’ small models to win the high-volume calls.
What it does not show: the figures are Meta’s own, several benchmarks carry an “-AA” variant suffix that makes them hard to compare with other labs’ numbers, and the post does not give a parameter count or say whether the weights will be released.