By Hamza Faisal
Meta's Superintelligence Labs has launched Muse Ember, its smallest and most efficient AI model to date, built for the repetitive work AI agents do all day. Announced on Monday, November 9, the model is available now through the Meta Model API and Muse Code, Meta's coding agent, under the model name muse-ember-1.0.
Ember is designed for the unglamorous layer of agent work: tool calls, coding subtasks, routing requests, extracting data and parsing long documents. Meta's pitch is that agents burn most of their tokens on exactly these small, structured steps, so a cheaper model that handles them well matters more than another flagship.
Haiku-level scores, a third of the cost
Meta measures Ember on the Artificial Analysis Intelligence Index, where it scores 36 at its highest reasoning setting. That matches Anthropic's Claude Haiku 5.5 and its best score, but at about $0.14 per task versus roughly $0.41 for Haiku. Meta says Ember beats OpenAI's GPT-6 Luna at every reasoning setting, and that no model on the full Artificial Analysis leaderboard scores higher for less money per task.
The price list is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. For comparison, Meta's own larger Muse Spark 1.3 scores 37 on the same index, one point higher, but costs about $2.39 per task, roughly 17 times Ember's price.
Ember handles text, images, audio and video, takes in up to one million tokens at once, and can return up to 512,000 tokens in a single response. It also supports the Responses API, and Meta lists it on OpenRouter.
Built for the subagent era
Meta describes Ember as a subagent model, the smaller model a bigger planner hands routine jobs to. It is the fourth model in the Muse family, following the mid-weight Spark in April, its open-source twin Glimmer in August, and the frontier Wonder model in October. Meta says Wonder plans the hardest problems, Spark handles coding and agentic workflows, and Ember does the repeatable versions of those tasks at scale. Ember is also the first model Meta says was trained using Muse Wonder Full.
There are limits. On FrontierCode 2.0-AA, a real coding benchmark, Haiku 5.5 still leads Ember, 34.9 percent to 27.2 percent, a reminder that efficiency scores do not guarantee superiority on every engineering task. Meta is also keeping Ember's weights private, as it did with earlier Muse releases, and says it evaluated the model under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks.
Why small models are the new battleground
The launch lands as the AI industry's centre of gravity shifts from flagship demos to the economics of running agents at scale. If agents are going to handle thousands of small steps per task, the cost of each step decides whether the product works as a business, and Meta, Anthropic and OpenAI are now competing as much on cost per task as on raw capability. Ember's arrival also continues a busy stretch for Meta's AI lab: its consumer Muse Agent launched in September, with new features including Spaces and video chat rolling out earlier this month.
