Meta releases Muse Ember, a small, low-cost AI model built for the repetitive steps inside AI agents
Meta says Ember matches a rival Anthropic model’s best score at about a third of the cost. It can read screenshots, scanned papers, voice memos and video.
Meta released Muse Ember, which it calls its smallest and most efficient AI model, on Nov. 9, according to a post on Meta’s AI research blog. Developers can use it now through the Meta Model API, a paid service for building Meta’s AI into other apps, and in all plans of Muse Code, Meta’s coding subscription.
What it is for
An AI agent is a program that carries out a multistep job on its own, such as handling customer complaints or updating records across several apps. Meta says agents spend most of their effort on small, repeated steps: calling other software tools, pulling facts out of documents, deciding where to send a task and doing small coding chores. Ember is built for that work.
Meta suggests pairing it with a bigger, pricier model. Muse Wonder, Meta’s model for the hardest problems, would plan the job, and many copies of Ember would do the volume work, Meta’s developer blog says. Ember can take in text, images, audio and video, so it can read a screenshot, a chart or a scanned page, or work from a voice memo or meeting recording without a separate model.
Price and Meta’s claims
Developers pay 10 cents per million tokens sent in and 50 cents per million tokens the model writes. A token is a chunk of text, roughly three-quarters of an English word, according to OpenAI’s help center. Ember can take in up to 1 million tokens at once, enough for a full set of computer code or documents in one request, Meta says.
Meta says Ember matches the best score of Anthropic’s Claude Haiku 5.5 at about a third of the cost and beats every setting of OpenAI’s GPT-6 Luna, using rankings from Artificial Analysis, an independent firm that tests AI models. Those are Meta’s comparisons. Meta says the price is “low enough to run in every product, for everyone.”
Meta says it tested Ember before release for chemical, biological, cybersecurity and loss-of-control risks and found it below its high-risk thresholds. It also checked whether Ember resists hidden instructions in web pages or documents that try to redirect it.
Source: Meta AI research blog, Nov. 9, 2026 and Meta developer blog.