Meta Muse Ember: An agent task costs 14 cents. Your process doesn't

How many times has your team shelved an automation because the API bill for thousands of runs looked unpredictable? That argument is collapsing. Not because of a new frontier model, but because of a small one.

What's new

• The model: Meta calls Muse Ember the smallest and most efficient model in the Muse family, built for repeatable agent work: tool calls, routing, extraction, coding subtasks, long documents and screenshots.
• The price: $0.10 per million input tokens and $0.50 per million output tokens via the Meta Model API. Prompts over 200K tokens cost three times as much.
• The performance: On the Artificial Analysis Intelligence Index, Ember scores 36, Muse Spark 1.3 scores 37. Per index task, Meta lists about $0.14 for Ember and $2.39 for Spark.
• Availability: since November 9 via Muse Code and the Meta Model API, plus Google Cloud and Oracle Cloud. Context window: 1 million tokens.

Price per token is the wrong number

Here's what I find interesting: Meta isn't selling Ember on token price but on cost per finished task. One point lower on the index, roughly a seventeenth of the cost. For marketing, that means enriching product feeds, sorting form leads, extracting data from invoices and briefs, tagging campaigns. All the jobs that used to stall on cost are now pocket change.

And Meta itself? The company that serves your ads is also becoming the supplier of your agents. With Box, Cloudflare and Oracle as partners, and on Google Cloud. That's a platform move, not a research note.

Cheap scales mistakes too

My rule for a first agent project: one clearly scoped job that ends in a concrete decision and produces a result someone can judge. A cheap model doesn't change that. It makes the rule matter more.

Because AI doesn't fix broken ways of working, it widens the cracks in the foundation. When a run costs next to nothing, you'll soon have ten thousand of them. With messy feeds and no approval step, that's ten thousand wrong product descriptions, not ten thousand saved hours. From now on, the expensive part is the work before: clean data, a clear task, someone who checks the output.

The catch

The benchmarks come from Meta and Artificial Analysis; independent real-world tests are still missing. On some coding tests, Meta's own table shows Claude Haiku 5.5 ahead. The announcement doesn't say whether or how Ember is offered in the EU. And as always: read the terms before you scale.

What I'd do now

• Pick one routine. The one job your team does by hand every week, like feed maintenance or lead sorting. One job, not five.
• Test with your own material. Run 50 real cases through a small model and compare against a large one. Count errors, not tokens.
• Plan the approval. Who spot-checks before anything goes live? Without that person, it stays an experiment.

Bottom line: When AI costs almost nothing, your cleanest process is the only competitive edge left.