Meta introduces Muse Ember, a small, low-cost model for repeated agent tasks

Meta announced Muse Ember on 9 November 2026 in a post on its AI research blog, calling it the smallest of its models and the most efficient. Meta says the model is built for the steps agents repeat throughout the day, such as routing, tool calls, coding subtasks, extraction and parsing long documents.

Meta says the model is out now on the Meta Model API and in Muse Code, and is also offered through Google Cloud and Oracle Cloud. It takes up to 1 million tokens of context, and Meta says it is the first model it has trained using Muse Wonder Full.

Meta lists prices of $0.10 per million input tokens and $0.50 per million output tokens for prompts of up to 200,000 tokens.

Meta says Muse Ember gives more intelligence for each dollar spent than Claude Haiku 5.5 or GPT-6 Luna at every reasoning setting, based on version 5 of the Artificial Analysis Intelligence Index, and that it reaches Haiku's best score at roughly one-third the cost. In Meta's own table, Muse Ember leads both models on AutomationBench-AA and Terminal-Bench 4.0, while Haiku 5.5 leads on FrontierCode 2.0-AA.

Meta positions Muse Wonder for its hardest problems, Muse Spark for coding and agent workflows, and Muse Ember for high volumes of routine coding and agent tasks. It says it tested Muse Ember before release under its Meta Superintelligence Scaling Framework and found it under the framework's high-risk thresholds in every category it assessed.