On this page · 5 sections ▾
Meta released Muse Ember on November 9, 2026, describing it as its smallest and most cost-efficient AI model. According to TechCrunch, it is built for work that agents repeat constantly, such as tool calls, coding subtasks, routing, extraction and long document parsing. It is available through the Meta Model API and Muse Code.
Key facts
- Price: $0.10 per million input tokens and $0.50 per million output tokens, plus $0.004 per million cached input tokens, per FoneArena.
- Context: 1 million tokens of input and up to 512,000 output tokens. It handles text, images, audio and video natively.
- Access: the Meta Model API and Muse Code as muse-ember-1.0. Muse Ember, Muse Spark and Muse Wonder are also on Google Cloud and Oracle Cloud.
- Benchmark claim: an Artificial Analysis Intelligence Index score of 36, the same as Claude Haiku 5.5 at its highest setting. GPT-6 Luna scores 28.
- Weights: kept private, as with earlier Muse releases, according to The Next Web, which cites The New Stack.
Cost per task is the pitch
Meta's central argument is intelligence per dollar. The Next Web reports that list prices match Haiku 5.5 and Luna, so the gap comes from cost per task on the Artificial Analysis index: about $0.14 for Ember, $0.41 for Haiku and $0.15 for Luna.
Digital Trends adds a comparison inside Meta's own lineup. Muse Spark 1.3 scores one point higher, 37, but costs $2.39 per task. On AutomationBench, which tests business workflows, Ember scored 37.8%, against 20.2% for Luna and 13.9% for Haiku.
Where it does not lead
The results are not uniform. On FrontierCode 2.0, which checks whether maintainers would merge a model's code changes, Haiku scored 34.9% and Ember 27.2%, per The Next Web. Digital Trends likewise notes Haiku still wins on real-world engineering tasks.
The Next Web also points out that several results, including Meta's Personal Superintelligence Index, come from Meta's internal tests. India Today reports Ember at 40.9% on Terminal-Bench 4.0, against 39.2% for Haiku and 16.4% for Luna. These figures were shared by Meta, and the sources do not describe independent replication beyond the Artificial Analysis index.
How Meta wants it used
Meta positions Ember as the executor under a larger planner. "Muse Wonder plans, Muse Ember handles the volume," the company wrote. TechCrunch gives the example of a customer-support agent: a heavier model sets rules for an issue, then a lighter one applies them across tens of thousands of cases.
Ember is the first model trained using Muse Wonder Full. Meta also said it ran Ember through its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and that it stayed below the high-risk thresholds in each category.
What to watch
TechCrunch lists Ember as the fourth Muse model, after Spark in April, Glimmer in August and Wonder in October. Meta said more efficient models are in training.
