Meta releases Muse Ember, an efficiency model for repetitive agent tasks

Meta's release of Muse Ember fits a pattern of AI labs segmenting their model lineups by task type rather than offering a single general-purpose system. Meta explicitly proposes a division of labor: its most complex workflows go to Muse Wonder, AI agent workflows and coding go to the mid-tier Muse Spark, and high-volume repetitive tasks go to Muse Ember. This tiering is aimed at controlling cost, since Muse Ember's per-task cost on the Artificial Analysis Intelligence Index is $0.14, compared with $2.39 for Muse Spark 1.3 (max).

The model is also notable as the first Meta model trained using its top-tier Muse Wonder Full model, suggesting a distillation-like approach where a smaller model learns from a larger one. In benchmark terms, Muse Ember matches Claude Haiku 5.5 on the Artificial Analysis Intelligence Index with a score of 36, and it outperforms both Haiku 5.5 and OpenAI's GPT-6 Luna on agent task benchmarks such as AutomationBench-AA and on coding benchmarks including Terminal-Bench 4.0, where it scored 40.9. However, on FrontierCode 2.0-AA, which measures whether code meets maintainer-mergeable quality, Muse Ember scored 27.2, below Haiku 5.5's 34.9.

Meta's emphasis on cost is central: it claims Muse Ember is the lowest-cost among the three models compared on a per-task basis. The model's multimodal support and 1M-token context window also position it for processing large document sets, a common need in enterprise agent deployments. The pricing of $0.1 per 1M input tokens and $0.5 per 1M output tokens, with cached input at $0.004, is designed for workloads that run many small tasks repeatedly rather than one large query.