Meta has released Muse Ember, touted as its smallest and most efficient model. It is not designed to tackle the most complex reasoning tasks, but rather to handle the most frequent tasks in an Agent’s daily operations—continuously invoking tools, retrieving and extracting information, reading code, running tests, making localized file modifications, and processing long documents. The model supports a context length of 1 million tokens and natively understands text, images, audio, and video. Meta has drawn a clear division of labor: Muse Wonder tackles difficult problems, while Ember handles the bulk of routine work. What’s truly remarkable is the cost. Muse Ember’s standard API pricing is $0.10 per million input tokens and $0.50 per million output tokens. On the Artificial Analysis Intelligence Index, it scores 36—tied with Claude Haiku 5.5’s top tier and surpassing all GPT-6 Luna reasoning tiers. Yet, to complete a benchmark task, Ember averages just $0.14, while Haiku 5.5 requires $0.41. Interestingly, both models have identical listed prices for input and output tokens. So why is Ember roughly two-thirds cheaper? Artificial Analysis attributes this to Ember’s lower cost for cached inputs and its output token consumption being only about one-third that of Haiku. This is the real efficiency breakthrough of the Agent era: models must not only be capable, but also speak less and consume less—driving down the cost of every step in repeated executions. Even more intriguing, Ember is beginning to compete with Meta’s own larger models. In the same Artificial Analysis tests, Muse Spark 1.3 scores 37 on the intelligence index, while Ember scores 36—but their per-task costs are $2.39 and $0.14 respectively, a difference of about 17x. Of course, Ember doesn’t outperform across the board. In the FrontierCode 2.0 test—which measures the real-world mergeability of engineering code—it still lags behind Haiku 5.5. Meta has since deployed Ember, Wonder, and Spark 1.3 on both Google Cloud and Oracle Cloud. Google Cloud is currently in preview mode, and Ember currently supports only text and image inputs. I believe Meta’s true message with this release is that the competition among large models is shifting from a singular contest of raw intelligence to a refined division of labor. In the past, everyone debated which model was the smartest. In the future, developers may care more about which tasks warrant the expense of hiring the smartest model—and which can be handled just fine by a model 17 times cheaper. After all, an Agent’s true value isn’t determined by occasional brilliant reasoning, but by reliably executing hundreds of ordinary operations. The smartest model defines an Agent’s upper limit of capability—but the cheapest model that’s still smart enough may determine whether an Agent can become a profitable business.