Meta has launched Muse Ember, its smallest and cheapest AI model yet. It isn’t trying to solve the world’s hardest problems. Instead, Meta designed it for the small, repetitive tasks AI agents keep doing, like using tools, fixing small bits of code, sorting requests, grabbing data from files, and reading lengthy documents.
Mark Zuckerberg announced the launch on X, saying you get more brains for your money with Muse Ember than with similar small models.
How cheap is Muse Ember?
It’s a lot cheaper than you’d expect from a model this capable. You pay $0.10 for every million tokens you send in and $0.50 for every million tokens it sends back. According to Meta, Claude Haiku 5.5 needs to spend about three times as much to hit the same top score, and GPT-6 Luna falls behind no matter which setting you pick.
On Artificial Analysis, Muse Ember finishes a task for roughly $0.14. Meta’s own Muse Spark 1.3 scores just one point higher but costs $2.39 for the same task.
Can a small model really keep up?
Based on the benchmark results Meta shared, it does a pretty good job. On AutomationBench, which tests business workflows, Muse Ember scored 37.8%, nearly three times Claude Haiku 5.5’s 13.9%. It also comes out ahead of both rivals at operating a computer and spotting visual details. That said, it isn’t perfect, and Claude Haiku 5.5 still wins on real-world engineering tasks.
Muse Ember also works with more than just text. It can look at images, listen to audio, and watch video too. Thanks to its huge 1M-token context window, you can feed it an entire coding project or a pile of files at once.
You can try Muse Ember today through the Meta Model API, Muse Code, Google Cloud, and Oracle Cloud. If you are building AI agents on a budget, Muse Ember is definitely worth a look.
