Meta is expanding its Muse series of models with its cheapest and most cost-effective variant to date. Muse Ember competes with Anthropic’s Claude Haiku 5.5 and achieves comparable scores in Meta’s own tests. Nevertheless, it is three times cheaper based on API calls.
Ember has been available since yesterday via the Meta Model API and Muse Code. The model has a conventional context window with a limit of 1 million tokens and generates up to 512,000 tokens per output. Meta states that this is sufficient for an entire codebase or a set of documents in a single call.
As mentioned, Muse Ember is, above all, inexpensive. Each million tokens of input costs 10 cents, and for outputs, users pay $0.50 per million tokens. Cached inputs, which make up the bulk of the AI context during long-running sessions, cost just $0.0004 per million tokens. However, there is a surcharge for exceeding 200,000 tokens; all rates then triple.
Ember is the first model that Meta has trained using Muse Wonder Full, which is a distillation of a larger model. This expands the line that began this year with Muse Spark, the first model from Meta Superintelligence Labs.
Haiku-level performance at an affordable price
On Artificial Analysis’s Intelligence Index, Muse Ember scores 36 points. That’s the same as Claude Haiku 5.5 at the maximum reasoning setting. However, a task costs about $0.14 with Ember, compared to $0.41 with Haiku. GPT-6 Luna doesn’t go beyond 28 points. While these scores don’t tell the whole story, they are generally indicative. The cost advantage stems primarily from less generated output and cheaper cache processing.
The biggest difference lies in AutomationBench-AA, which tests business workflows from start to finish. Ember achieves 37.8 percent, while Haiku achieves only 13.9 percent. Still, the model doesn’t win across the board: on FrontierCode 2.0-AA, Haiku scores 34.9 percent compared to 27.2 percent for Ember. Additionally, some of the figures come from Meta’s internal evaluations.
The price also stands out even within its own family. Muse Spark 1.3 scores 37 points on the index but costs $2.39 per task. The model released in September thus appears to have already been surpassed on all fronts within Meta’s own lineup.
Task allocation within Muse
Meta primarily views Ember as a complement to the more powerful models. Muse Wonder, for example, handles planning and directs Ember agents to specific tasks to achieve the goal. Box has already tested the model. “Muse Ember is a fast, efficient foundation for agents that actually perform work on business content,” says Yashodha Bhavnani, VP of AI Products at the company.
Ember, Spark, and Wonder are now also available via Google Cloud and Oracle Cloud. Meta says it has even more efficient models currently in training.