A small model for agents' daily work
Meta has announced Muse Ember, which the company describes as its smallest and most efficient artificial intelligence model to date. According to The Decoder, the model is optimized for the routine tasks that agents perform throughout the day: calling external tools, handling small coding tasks, routing requests, extracting data, and processing long documents.
The decision to focus on these tasks signals a clear direction: while frontier models are presented primarily for their general reasoning capabilities, Muse Ember focuses on efficiency in the repetitive operations that agents perform at high volume. According to the source, the model handles text, images, audio, and video and can process up to one million tokens at once.
The cost comparison and coding limitations
Meta claims that Muse Ember offers more intelligence per dollar, at every reasoning level, than Claude Haiku 5.5 and GPT-6 Luna. This is a company statement, reported by the source but not independently verified in the available material. The cited cost figure relates to the Artificial Analysis test: about $0.14 per task for Muse Ember, versus $2.39 for the larger Muse Spark 1.3, which achieves an almost identical score, 37 versus 36.
The picture changes, however, for real-world programming tasks. In the FrontierCode 2.0-AA tests, according to The Decoder, Haiku beats Muse Ember by 34.9 percent to 27.2 percent. Meta's model leads in most of the company's internal tests, but the coding result shows that optimization for repetitive tasks entails trade-offs in other capabilities.
Pricing and availability
Muse Ember is available immediately through the Meta Model API and Muse Code, at a price of $0.10 per million input tokens and $0.50 per million output tokens, again according to the source. The combination of low cost per token and the claimed efficiency in routing and data extraction tasks is the central argument of the presentation.
The real extent of the cost advantage over the cited competitors and the model's performance under workloads other than those indicated by Meta remain to be independently verified. The material does not report third-party results or details about architectures or training times, elements that would be necessary for a complete assessment.
