Meta Releases Muse Ember: A Model 17x Cheaper Than Muse Spark With Nearly the Same Score

Most of an AI agent's work is boring. It calls APIs, decides which tool to hand a request to, pulls fields out of an email, and skims through a long PDF. None of this requires the smartest model on the market, yet many teams pay for one anyway because it is simpler. Meta is betting that this will soon stop being the norm, and it is releasing Muse Ember.

What the model is

Muse Ember is Meta's smallest and most efficient model to date. It is aimed not at chatbots but at the workhorses of agentic systems: tool calling, small coding tasks, request routing, data extraction, and long-document processing. The model is multimodal: it accepts text, images, audio, and video. The context window is up to a million tokens.

It is available now through the Meta Model API and Muse Code. The price is $0.10 per million input tokens and $0.50 per million output tokens. As a rough guide, a 100,000-token document will cost about one cent on input. Output is more expensive, but for routine tasks it is usually short: JSON, a class label, a function name.

The numbers: what impresses and what to verify

Meta's main argument is "intelligence per dollar." According to the company, at every reasoning level Ember beats Claude Haiku 5.5 and GPT-6 Luna on this metric. The most telling example is the comparison with Meta's own flagship. In the Artificial Analysis test, Ember scores 36 points on the intelligence index, while Muse Spark 1.3 scores 37. Yet a task on Ember costs about $0.14, versus $2.39 on Spark. That is a nearly 17x difference for a single index point.

There are caveats. Leadership in most tests comes from Meta's own benchmarks, and such measurements are traditionally chosen to play to a model's strengths. Independent Artificial Analysis confirms the price and index figures, but that does not remove the need to run your own data.

The main weakness is real code. On FrontierCode 2.0-AA, Haiku scores 34.9% and Ember 27.2%. That is a noticeable gap: the model is good at small edits and function calls, but don't expect it to replace a strong model on complex programming tasks.

What to do in practice

The most sensible scenario is a cascade. The cheap model takes on the stream of repetitive operations: classifying incoming requests, parsing documents, calling tools, filling in structures. The expensive one steps in only when the task is truly hard, such as refactoring a large repository. If your agent's entire chain currently runs on a flagship, it is worth estimating what share of calls could move to Ember without a loss in quality.

What to check first:

• tool-calling accuracy on your own function schemas, not on public benchmarks;
• robustness on long documents: a million tokens in the spec does not guarantee that the model won't lose details in the middle;
• cost on a real task, not per token: a reasoning model can spend more output tokens than the price list suggests;
• behavior on multimodal inputs, if your pipeline includes audio or video.

One more point: Ember is available only through Meta's ecosystem, the Model API and Muse Code. For teams already tied to other providers, that means extra integration work and vendor lock-in concerns.

The bottom line is simple. The small-model market is shifting from "how smart is it" to "how much does one completed task cost." Muse Ember is a strong argument for that shift, but choose it based on your own measurements: for routine work it looks very attractive, while for serious code it still trails Haiku.