Meta releases Muse Ember: a benchmark task costs $0.14 versus $0.41 for Haiku

If your agent parses hundreds of invoices, looks up CRM records and moves data between services, a substantial share of its calls involves repetitive actions. Paying for the capabilities of the most expensive model at every step can get costly. On November 9, Meta introduced Muse Ember, a compact worker model for these tasks. The practical question is simple: can it complete your workflow at a lower cost and with an acceptable number of errors?

What Meta proposes handing over to Ember

Meta's documentation lists tool calling, routing, data extraction and processing large collections of documents. The company proposes a pairing: Muse Spark or Muse Wonder creates the plan, while Ember handles bulk operations. The model accepts text, images, video, audio and PDFs, and returns text. This allows it to handle different types of source material within a single workflow. Meta documentation.

Consider an incoming invoice. The system needs to identify the supplier, extract the amount and date, match the purchase order and prepare an accounting entry. The tools and required fields can be defined in advance. A complicated case, such as an amount that does not match the purchase order, goes to a more capable model or a human.

This division of labor is worth testing on tasks with clearly defined outcomes. Did the system identify the right supplier? Did the amount match? Were all required fields completed? The clearer the criteria, the easier it is to determine whether a cheaper worker actually saves money. It also reveals which exceptions need a different route.

Where the promised savings come from

The source material cites a comparison published by Meta: about $0.14 per task for Ember versus $0.41 for Claude Haiku 5.5. That is nearly a threefold difference. But the cost of a benchmark task cannot automatically be applied to processing your invoices or support requests. Meta specifies that the comparison uses Intelligence Index v5, the models' list prices and the maximum reasoning level. Comparison conditions.

Ember's base price is $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs $0.004. For requests longer than 200,000 tokens, all three rates triple. The stated limits are a one-million-token context window and up to 512,000 output tokens. Meta announcement.

A large context window is useful when access to an extensive set of documents is needed. But it does not, by itself, establish that loading the entire archive into every request is cost-effective. For a pilot, it is useful to compare two approaches: the full collection and a small set of relevant documents. Then calculate the cost of a correct result, including retries and corrections.

Where the promise ends and testing begins

In the published AutomationBench-AA table, Ember scores 37.8%, GPT-6 Luna 20.2% and Haiku 5.5 13.9%. That is a strong showing for automation, but even the best score here is far from error-free performance. Some other results in the announcement are labeled as Meta's internal evaluations. These measurements help identify a candidate for a pilot; the reliability of a specific workflow still needs to be measured separately. Results and methodology.

Developers should take a few dozen typical tasks and add difficult exceptions: incomplete documents, conflicting fields and an unavailable tool. Each task needs an expected outcome and a set of permitted actions. Compare models using the same tools and validation rules, recording latency, retries and human interventions.

For a product manager, the overall economics of the workflow matter more. A cheap call can prove expensive if an employee then has to repair incorrectly entered data. The pilot's primary metric should therefore be the cost per successfully completed operation. Ember is particularly worth testing where there is a lot of repetitive work, results are easy to verify and errors can be caught before data is written to the production system.