Meta unveils Muse Ember, its smallest model for cheaper AI agents

Meta has unveiled Muse Ember, which it presents as its smallest and most efficient AI model so far. According to The Decoder and Hipertextual, both reporting on Meta's announcement, the model is built for the routine work that AI agents carry out in the background, not for open-ended conversation or heavy reasoning.

Muse Ember is a small language model, meaning a model deliberately kept compact so that each request is cheaper to process. Meta says it targets calling tools, completing small programming tasks, routing requests, extracting data from files and analysing long documents. The company argues that these repetitive steps consume most of the resources when an AI agent works on its own, which is why a dedicated low-cost model makes sense.

What Meta announced about Muse Ember

Ember joins the Muse family developed by Meta's superintelligence team, alongside Spark, Glimmer, Wonder and other members, according to Hipertextual. The model is available immediately through the Meta Model API and through Muse Code, Meta's coding tool. The two outlets report the following figures, all of which originate from Meta or from benchmarks that Meta cites:

• Price: 0.10 dollars per million input tokens and 0.50 dollars per million output tokens, according to both The Decoder and Hipertextual.
• Context window: up to one million tokens of input at once, and up to 512,000 tokens of output according to Hipertextual, enough to process a whole codebase or a collection of documents in a single request.
• Modalities: native understanding of text, images, audio and video, so an agent can analyse screenshots, diagrams, scanned documents or voice recordings without calling a second model.
• Cost per task: about 0.14 dollars in the Artificial Analysis test, according to The Decoder.

A token is the unit of text a model reads or writes, roughly a fragment of a word, and pricing per million tokens is the usual way AI vendors bill access to their models through an API. Those prices put Ember among the cheapest options for high-volume agent workloads, although the two articles do not compare them with a competitor's published rates.

How Muse Ember compares with Claude Haiku 5.5 and GPT-6 Luna

According to Meta, as reported by The Decoder, Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at every reasoning level. In the Artificial Analysis test, it costs about 0.14 dollars per task and is the cheapest model to reach an Intelligence Index score of 36 or higher. Claude Haiku 5.5 reaches the same score at 0.41 dollars per task, almost three times as much. The larger Muse Spark 1.3 scores 37, almost the same, but at 2.39 dollars per task.

On AutomationBench-AA, a benchmark of business workflows, Muse Ember scores 68.1 percent, on a par with Muse Wonder and GPT-6 Astra, The Decoder reports. Hipertextual adds that Meta says Ember matches Claude Haiku 5.5 on some benchmarks and beats GPT-6 Luna on others, always at a lower cost per task. On OSWorld, which measures how well an agent can control a computer on its own, Ember beats GPT-6 Luna and Claude Haiku 5.5, all run at maximum reasoning. On Terminal Bench 4.0, which tests real engineering tasks, Ember scores more than twice as high as GPT-6 Sol, but Hipertextual notes that the top of the ranking belongs to Anthropic models (Opus, Astra and Sonnet) while Ember competes in the middle with other large open-source models.

A division of labour between small and large models

Meta describes a two-tier approach. Advanced work such as planning an overall strategy can go to Muse Wonder, while Ember takes care of operational work. According to Hipertextual, Meta's engineers tuned the model for programming subtasks such as reading files, running tests and making targeted fixes, which are often repeated hundreds of times in a single development session. At that frequency, a difference of a few cents per task adds up quickly.

On safety, Hipertextual, citing a post on Meta's blog, says the model has the same training-time and run-time safeguards as the other Muse models, meaning it has no hacking capabilities or other prohibited abilities. The outlet still advises care with the permissions you grant an AI that works on your files.

What remains unverified

Every performance figure above comes from Meta or from benchmarks it cites, and the two articles we read do not mention any independent replication. The per-task cost reported by Artificial Analysis combines the price per token with the number of tokens a model actually consumes, so it can differ from one workload to another. Neither article gives the number of parameters of Muse Ember or says whether its weights are open. Treat the comparisons with Claude Haiku 5.5 and GPT-6 Luna as claims from the vendor until third parties publish their own measurements.

What it changes for companies running AI agents

For a business that deploys agents, the practical lesson is about routing: reserve the expensive models for planning and send repetitive steps such as extraction, triage and small code fixes to a cheaper one. Muse Ember makes that choice easier to test, since its published price is 0.10 and 0.50 dollars per million tokens. Before switching, measure the cost and success rate per task on your own documents and code, because Meta's benchmarks may not reflect your workload. It also pays to set a budget per agent, so that cheap tokens do not turn into an uncontrolled bill.