An agent that searches a database, calls a tool, extracts fields from a document or completes a small coding task may perform useful work many times before a person sees the result. That changes the economics of AI. A model does not need to be brilliant at every step, but it does need to be dependable enough that its errors do not erase the savings.
Meta is targeting that problem with Muse Ember, its smallest and most efficient model for repetitive agent tasks. In Meta’s launch announcement for Muse Ember, the company says the model is priced at $0.10 per million uncached input tokens and $0.50 per million output tokens. It is intended to execute subtasks planned by larger Muse models rather than replace them.
The proposed arrangement resembles a manager assigning routine work to a large operations team. A stronger model, such as Muse Wonder, decides what should happen. Ember then carries out the individual steps at scale. Meta’s Muse Ember API documentation describes this planner-plus-worker design and says Ember can work with Muse Spark or Muse Wonder.
That division could make agent systems more affordable. A powerful model would not need to read every document, classify every message or make every tool call. Ember also supports text, images and audio, according to the documentation, and has a context window of 1,048,576 tokens with a maximum output of 524,288 tokens. Those capabilities make it suitable for large files and multimodal workflows, at least on paper.
The more difficult question is whether the smaller model is cheap enough after accounting for supervision. A low-cost worker that occasionally corrupts a record, misroutes a request or writes faulty code can force a larger model, or a human, to review its work. In that case, the apparent bargain becomes a form of deferred expense.
Artificial Analysis reports that Ember scores 36 on its Intelligence Index and estimates its cost at $0.14 per index task. Its analysis says that places Ember at a similar intelligence score to Claude Haiku 5.5 while costing substantially less on the measured task. The comparison is useful, but it is not the same as proving that Ember will reduce the total cost of a production agent.
The criticism is centered on that distinction. Ridge argued in its assessment of Ember that Meta’s cost-efficiency case is compelling but that the company’s comparisons deserve caution. Several results reportedly come from Meta’s own testing, coding harnesses can produce different outcomes, and Ember is a closed model with limited independent hands-on evidence at launch.
That concern matters because agent performance is unusually sensitive to workflow design. A model may look strong in a short benchmark but struggle when it must preserve state across many tool calls, recover from an ambiguous instruction or recognize that an earlier step failed. Repetition magnifies small weaknesses. If an agent completes 1,000 routine actions, even a modest error rate can create dozens of cases requiring investigation.
The review from eesel is more positive about practical uses such as routing, tagging, extraction and sub-agent work. But it also points to a weaker FrontierCode result, the absence of open weights, no batch discount and higher pricing for long-context work. Those objections go directly to buyers who care less about a headline token price than about predictable costs across a real workload.
Artificial Analysis also flags weaker factual-knowledge performance and mixed coding results in its independent comparison. That combination suggests Ember may be most useful when the task is narrow and easy to verify, rather than when it must make consequential judgments on its own.
The practical test, then, is not whether Ember can match a larger model on a leaderboard. It is whether a planner-worker system can divide labor without multiplying failure points. Companies should measure completed tasks, review time, retries and error recovery alongside token consumption.
Meta has made the worker cheap. The market still has to determine whether it is reliable enough to trust with the office’s repetitive work.
- InvadingInvader · CC BY-SA 4.0
Daniel Reyes is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Daniel Reyes's usual length and in Daniel Reyes's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.