Meta Announces Muse Ember, a New AI Model for Complex Tasks

Meta Superintelligence Lab has introduced Muse Ember, the company's lightest and most cost-efficient artificial intelligence model to date.

According to the announcement made by Meta, the new model was specifically designed for repetitive agent jobs such as complex reasoning execution, tool calls, coding subtasks, routing, extraction, and long document parsing. Based on data derived from Artificial Analysis composite benchmark scores, Ember is positioned as both a cheaper and smarter alternative to Anthropic's Haiku and OpenAI's Luna models, with a price of 10 cents per million input tokens. It can also compete cost-wise with open-source competitors such as DeepSeek V4.1 Flash and GLM 5.3 Flash.

The model is structured to manage high-volume tasks by pairing with a system such as Meta Muse Wonder, where larger rules are established. Offering a multimodal structure capable of processing text, images, and audio, Ember can work with screenshots, charts, scanned documents, voice notes, meeting recordings, and videos without the need for separate models. Launched as the fourth model in the Muse series following Spark, Glimmer, and Wonder, Ember aims to reduce the cost of high-volume operations in agent-based tasks.