Meta announced Muse Ember, its smallest and most efficient artificial intelligence model to date. Unveiled by Mark Zuckerberg in a post on X, the new model was developed to perform repetitive tasks at scale, especially those carried out continuously by AI agents, such as tool calls, document analysis, data extraction, and programming subtasks.
The model functions as alightweight agent or subagentwithin more complex artificial intelligence systems. Developed from Muse Wonder Full, the company’s most advanced model, Ember prioritizes speed and low operating costs. It has a context window of 1 million tokens and can generate up to 512,000 output tokens per call. It also processes text, images, audio, and video without relying on auxiliary models.
Muse Embersurpassed Claude Haiku 5.5 and GPT-6 Luna in evaluations of enterprise automation, complex agent tasks, terminal programming, and visual perception. In Artificial Analysis’s overall intelligence index, the model scored 36 points, tying with Claude Haiku 5.5 and surpassing GPT-6 Luna’s 28 points.In FrontierCode 2.0-AA, which evaluates real-world software engineering tasks, Claude took the lead with 34.9%, compared with 27.2% for Ember.
Price is one of the main attractions of the launch. On Meta’s API, the model costs US$0.10 per million input tokens and US$0.50 per million output tokens for prompts of up to 200,000 tokens. In Artificial Analysis evaluations, the average cost per task was approximately US$0.14. This amount is about 17 times lower than that recorded by Meta’s own Muse Spark 1.3, which costs US$2.39 per task.
Meta also evaluated the model under the Meta Superintelligence Scaling Framework, which analyzes risks related to cybersecurity, chemical and biological threats, and loss of control over AI systems.According to the company, Muse Ember fell below all high-risk thresholds in these categories.
The model also received safeguards for tasks performed on behalf of users, including resistance to malicious instructions hidden in pages and documents (prompt injection) and requiring human confirmation before high-impact actions.
With the launch, Meta completes its portfolio of Muse models: Muse Wonder, for more complex problems; Muse Spark, for workflows and programming; and Muse Ember, for repetitive tasks and large-scale agent execution.
Who can use Muse Ember?
The model primarily servesdevelopers, engineering teams, and businesses that need to run AI agents at scale. The aim is to reduce spending on recurring tasks that consume many tokens throughout the day.
Muse Ember can serve as the primary agent in automated processes or as a subagent responsible for specific activities within more complex systems.Applications include the analysis of lengthy documents, code review, information extraction, and the execution of commands in external tools.
The model is available through the Meta Model API, with the identifier muse-ember-1.0, in Muse Code, and on partner cloud platforms, including Google Cloud and Oracle Cloud.
If you enjoyed the content, you may also want to check out “Can you disable Meta AI on WhatsApp? No, but you can use these tricks”.
