Meta released Muse Ember on November 9, 2026, describing it on its research blog as "our smallest and most efficient model." It is built for high-volume agent work such as tool calls, coding subtasks, routing, data extraction and long-document parsing. Meta says it is the first model trained using Muse Wonder Full.
Muse Ember is available now on Muse Code and the Meta Model API under the ID muse-ember-1.0. Meta lists pricing of $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. The model has a 1 million token context window, supports up to 512K output tokens, and accepts text, images, audio and video. It supports the Responses API. Meta did not disclose a parameter count or a license.
Meta's reported benchmarks compare Ember with Claude Haiku 5.5 and GPT-6 Luna at maximum reasoning. Ember scored 37.8% on AutomationBench-AA (Haiku 5.5: 13.9%; GPT-6 Luna: 20.2%), 40.9% on Terminal-Bench 4.0 (Haiku 5.5: 39.2%), and 69.5% partial credit on OSWorld 2.1 (Haiku 5.5: 68.9%). It trails Haiku 5.5 on FrontierCode 2.0-AA, scoring 27.2% against 34.9%.
On the Artificial Analysis Index, Ember scored 36, tying Haiku 5.5. Meta says it "matches Haiku's best score at about a third of the cost." Meta puts the cost per Artificial Analysis task at about $0.14, compared with $2.39 for its larger Muse Spark 1.3.
Meta says it evaluated the model under its Meta Superintelligence Scaling Framework and found it remained below the framework's high-risk thresholds in the categories assessed.
For teams running large agent pipelines, Ember is a low-cost option for sub-agent and routing steps that need long context and multimodal input. Meta's own benchmarks suggest it is weaker than Haiku 5.5 on harder coding tasks.