Muse Ember pricing: 14 cents a task vs Spark’s $2.39, explained

Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens through the Meta Model API. A token is a small chunk of text, often part of a word, and AI services bill by how many of them a model reads and writes.

Here is the catch that makes per-task cost the better measure. Two models with identical token prices can burn very different amounts of tokens to reach the same answer, especially at higher reasoning settings where a model “thinks” longer before replying. Meta says Haiku 5.5 needs to spend about three times as much as Ember to hit the same top score, and that Luna trails Ember whichever setting you choose.

• Input: $0.10 per million tokens
• Output: $0.50 per million tokens
• Cost per benchmark task (Artificial Analysis): about $0.14 for Ember, $0.15 for Luna, $0.41 for Haiku 5.5

One small wrinkle: one early report described Ember’s input rate as far below Haiku’s and Luna’s, while Meta’s own comparison and several other reports put all three at the same list prices. The per-task figures above are the numbers Meta published, and they are where Ember’s saving shows up.

For most high-volume jobs, Meta’s own numbers suggest Ember is the better buy than Muse Spark 1.3. Spark 1.3 scored one point higher on the Artificial Analysis index, but its $2.39 per task makes it far more expensive for anything you run thousands of times.

The two models are aimed at different work. Meta positions Spark as its model for coding and other tool-heavy jobs, and the 1.3 update, released in September, cut tool calls by roughly 20% and token use by roughly 25% compared with the version before it. Ember is the cheap tier for steps that repeat: reading long documents, routing requests to the right place, pulling out data, handling small coding subtasks and making tool calls.

Then there is Muse Wonder, the frontier model Meta unveiled in October for its hardest reasoning problems. Meta’s suggested setup puts Wonder in charge of planning and hands the repetitive steps to Ember. In the company’s words, “Muse Wonder plans, Muse Ember handles the volume.” Ember is also the first model Meta trained with help from the larger Muse Wonder Full.

All three models share a context window of 1 million tokens and accept text, images, audio, video and PDFs.

Ember ties Haiku 5.5 on overall intelligence and beats Luna, but it does not win every test. On the Artificial Analysis Intelligence Index, Ember and Haiku 5.5 both scored 36 at their highest settings, while Luna scored 28.

Meta also published results from other agent and coding tests:

• AutomationBench (full business workflows): Ember 37.8%, Luna 20.2%, Haiku 13.9%.
• FrontierCode 2.0 (whether maintainers would merge a model’s code changes): Haiku 34.9%, Ember 27.2%.
• Terminal coding: Ember came out ahead of Haiku.
• Computer use and visual detail: Meta says Ember beat both rivals.

Read these with some care. Several of the results, including Meta’s Personal Superintelligence Index, come from Meta’s internal testing rather than an outside lab. The Artificial Analysis score is the independent one, and it shows a tie with Haiku rather than a clear win.

Ember’s entry price also undercuts several cheap Chinese models, including DeepSeek V4.1 Flash, MiniMax M3.1 Flash and Z.ai’s GLM-5.3 Flash, all of which score lower on the Artificial Analysis index.

No. Meta is keeping Ember’s model weights private, the same approach it has taken with its other Muse releases, so you can’t download it and run it on your own hardware. You call it as a hosted service and pay per token.

That marks a change from Meta’s earlier Llama models, which developers could download. The Muse family does include one open model: Glimmer, released in August as an open-source twin of Spark. Meta hasn’t announced an open version of Ember. It has said even cheaper models are already in training.

Ember is live now in four places:

1. The Meta Model API, Meta’s developer platform.
2. Muse Code, Meta’s coding tool.
3. Google Cloud, through managed APIs on the Gemini Enterprise Agent Platform, alongside Muse Spark 1.3 and Muse Wonder.
4. Oracle Cloud.

The Google Cloud listing is an odd twist, since Google builds Gemini, a direct rival to Muse. For companies, it means they can pay for Ember through a cloud contract they already have instead of signing a separate deal with Meta.

You won’t buy Ember directly as a consumer, but cheap models like it pay for the agent features you do use. When an AI assistant books a trip or sorts your files, it may repeat the same small step dozens of times. Each one costs the company tokens, and those costs decide which features get offered free, which go behind a subscription and which never ship.

Meta’s consumer push gives a sense of scale. Its Muse app, with the avatar named Jolly, passed 5 million downloads in 22 days after its September launch, and voice conversations and animated avatars began rolling out in the U.S. in early November. Meta has not said which model handles which steps inside the app.

One early business example shows the kind of work Ember targets. Box, the cloud storage company, said it used Ember in early testing to check contracts clause by clause against a company’s negotiation rules and give each contract an overall risk rating.