Meta releases Muse Ember for the repetitive tasks of AI agents

Meta Superintelligence Labs released Muse Ember on 9 November. It is a small AI model for the work that AI agents repeat all day: calling tools, pulling data out of documents, passing tasks on to the right part of a system and small programming jobs, such as reading files, running tests and fixing specific errors. An AI agent is an AI program that carries out a series of steps on its own to finish a task. Ember is available through the Meta Model API and the programming environment Muse Code, under the name muse-ember-1.0.

Much cheaper than Meta's larger models

Ember costs $0.10 per million input tokens and $0.50 per million output tokens. Tokens are the small pieces of text into which a model splits its input and output; a million tokens is roughly 750,000 words. Input that the model processed shortly before costs $0.004 per million tokens. For comparison: Meta's model Muse Spark 1.3 costs $1.25 for input and $4.25 for output, more than ten times as much.

Ember can read up to a million tokens in one go, enough for a whole collection of code or documents, and give an answer of up to 512,000 tokens. According to Meta, the model understands text, images, audio and video. This means an agent can work with screenshots, scanned documents and recordings of conversations without needing a separate model for them.

The large model plans, the small one does the work

Meta suggests combining Ember with its largest model, Muse Wonder: Wonder makes the plan, Ember does the bulk of the hands-on work. TechCrunch gives the example of a customer service agent: the large model sets out the rules for an issue and the small model applies them to tens of thousands of cases.

Ember is the fourth model under the Muse name, after Spark in April, the open-source model Glimmer in August and Wonder in October. According to Meta, Ember is the first model trained with the help of Muse Wonder Full.

To compare Ember with competitors, Meta uses the Intelligence Index of Artificial Analysis, a company that compares AI models with a fixed set of tests. According to Meta, Ember reaches the same score as Anthropic's Claude Haiku 5.5 at its strongest setting, for about a third of the cost per task, and scores higher than OpenAI's GPT-6 Luna at every setting. On a demanding programming test from the same set, however, Haiku scores clearly higher according to Meta's own figures. Meta also says that companies which have already tried Ember, such as Cloudflare and Box, are positive about it.

Meta tested Ember under its own safety framework, the Meta Superintelligence Scaling Framework. According to the company, the model stayed below the thresholds for high risk, including in the areas of cybersecurity and loss of control. Meta also tested whether Ember handles sensitive data carefully, ignores hidden instructions in web pages and documents, and asks a human for permission before taking a far-reaching step.

What does this mean for the road to the singularity?

An agent that finishes a large task often takes hundreds of small steps. If each step becomes much cheaper, AI can do much more work on its own for the same money. The division of labour that Meta proposes, a large model that thinks and a small model that carries out, shows how that can work in practice. How well Ember performs outside Meta's own tests will become clear as people use it in the coming weeks.

What is the technological singularity? Read the explanation →