Meta's new budget AI does routine office tasks for a third of the price
Meta released Muse Ember on 9 November, a small AI model built for the repetitive steps AI helpers run all day. Meta says it matches a leading rival's best score at about a third of the cost.
The short version
- Meta released Muse Ember on 9 November. It is the company's smallest and cheapest AI model.
- It is built for the small, repeated steps AI helpers run all day, such as pulling details from documents.
- Meta says it matches the best score of Anthropic's small model, Haiku, at about a third of the cost.
What Meta released
Muse Ember is made for the jobs an AI helper repeats many times in one session. Meta's list includes calling software tools and making small coding fixes. It also covers reading long documents and pulling out details.
It understands pictures and sound as well as text. So it can work from a screenshot or a meeting recording without a second model. It can also take in a whole document set in one go.
Companies can use it through Meta's paid developer service or two big cloud platforms. Meta charges by volume: $0.10 per million small pieces of text it reads, and $0.50 per million it writes.
How it compares
Most of Meta's figures come from Artificial Analysis, an independent firm that tests AI models. The rest are Meta's own tests.
- On a test of complete business workflows, Ember scored 37.8%. Anthropic's Haiku scored 13.9% and OpenAI's small model scored 20.2%.
- One standard task costs about $0.14 with Ember. Meta's larger Muse Spark scores about the same overall but costs $2.39 per task.
- It is not the best at everything. Ember is weaker at real software fixes: Haiku scored 34.9%, Ember 27.2%.
Early users point to office work. Box, a document software company, said Ember can score a contract against a company's negotiation rules clause by clause. It can also screen job candidates against a detailed brief.
Ember can also work a computer directly from screenshots, and read fine visual detail. On Meta's own test of using desktop apps, it scored about the same as Haiku and well ahead of OpenAI's small model.
Meta says it ran the same safety checks as on its larger models. That includes whether Ember resists hidden instructions in web pages or documents and asks a person before high-impact actions.
What it means for Indonesia
Meta says AI helpers spend most of their effort on many small, structured steps. Meta says that is where most of the bill comes from.
A common setup now pairs a large model that plans with a cheaper model that does most of the work. Meta pitches Ember for that second role, next to its bigger Muse Wonder.
Think of firms that handle customer service or document checks at scale. For them, the price per step decides whether a project pays back. Cheaper steps can change the math on a pilot that stalled on cost.
Box's example fits many Indonesian offices. Checking contracts against house rules, or screening job applicants against a brief, is repetitive work that eats senior staff time. A small model that does the first pass, with a person reviewing the result, is a sensible place to start.
Two cautions. These are Meta's launch figures, so test before you switch. And Meta's post says nothing about Bahasa Indonesia, so check quality in the languages your customers use.
What to do next
- List the AI steps your teams run most often and what each one costs today.
- Test Ember next to your current small model on real cases, in Bahasa Indonesia as well as English.
- Keep a larger model for planning and judgment calls, and measure where the cheap model falls short.
- Before sending customer records, check where each cloud provider processes the data.