Meta released Muse Ember on Monday, Nov. 9, the smallest and cheapest model in its Muse lineup. The company built it for the repetitive work AI agents do all day, such as tool calls and small coding jobs. Developers can use it now through Meta Model API as muse-ember-1.0, and Meta says it is included in every Muse Code subscription plan.
Meta is selling it on price. The company says Ember matches the best score of Anthropic's Claude Haiku 5.5 on an independent ranking for about a third of the cost per task.
What it costs
Ember costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens, according to Meta's model page. Longer prompts cost triple, $0.30 and $1.50. The model can read up to 1 million tokens in one request and write up to 512,000.
Meta also compared Ember with its own larger model, Muse Spark 1.3.
| Muse Ember | Muse Spark 1.3 | |
|---|---|---|
| Input, per million tokens | $0.10 | $1.25 |
| Output, per million tokens | $0.50 | $4.25 |
| Artificial Analysis Intelligence Index | 36 | 37 |
| Cost per index task | $0.14 | $2.39 |
Spark scores one point higher and costs about 17 times as much per task.
Individual programmers can get Ember through Muse Code, the coding tool Meta sells as a monthly subscription that runs in a computer terminal, with plans that start at $5 a month and go up to $200 for the heaviest users.
How it compares
Artificial Analysis, an independent group that tests AI models, gives Ember a 36 on its Intelligence Index. That ties Claude Haiku 5.5 at its highest setting. OpenAI's GPT-6 Luna scores 28. It puts the cost of running Ember through its test at $0.14 per task. It also lists Ember as proprietary, so developers cannot download the model and run it on their own machines.
Meta's wider benchmark table is more mixed. Ember leads Haiku and Luna on most of the agent and coding tests Meta published, including a business-workflow test where it scored 37.8% to Haiku's 13.9%. Haiku wins FrontierCode, which checks whether code changes would be good enough to merge, 34.9% to 27.2%. On long documents, Ember and Luna tie. Several of the results come from Meta's internal evaluations, and its cost-per-task chart is marked as estimates at proposed pricing.
What Meta says it is for
Meta wants Ember to be the worker under a bigger model. In the setup it describes, Muse Wonder plans a hard job and Ember carries out the many smaller steps. It handles text and images, plus audio and video. Meta says it can also operate a computer from screenshots.
It is the first Meta model trained with help from Muse Wonder Full, the company's most advanced model. Meta says safety tests under its own framework found Ember stayed below its high-risk thresholds in every category checked, from cybersecurity to loss of control.
Meta's announcement says nothing about bringing Ember to the Meta AI assistant in its consumer apps, though the company says even more efficient models are already in training.
