Meta unveils Muse Ember: 17 times cheaper than its own Spark, matching Claude Haiku 5.5

Beating AI Flash News: Meta has launched its lightest AI model, Muse Ember, designed to handle repetitive agent tasks including tool calls, code edits, and document analysis. Ember supports text, image, audio, and video inputs, with a 1 million-token context window. On Artificial Analysis’s Intelligence Index, Muse Ember scored 36 points, matching Claude Haiku 5.5, yet its per-task cost is just $0.14, roughly one-third of the latter’s. Compared to Meta’s own Muse Spark 1.3, Ember trails by only 1 point, but its cost plummets to around one-seventeenth. The model’s affordability stems from token utilization efficiency: while Ember and Haiku 5.5 have identical standard input and output per-token prices, Ember’s cache read cost is 60% lower, and it generates only about one-third of the output tokens Haiku produces for evaluation tasks. Meta plans to assign complex task planning to its large model Muse Wonder, while offloading repetitive work to Ember. For example, Wonder would draft a code migration plan, and multiple Ember agents would modify files and run tests in parallel. However, Ember has clear drawbacks: in Artificial Analysis’s knowledge test, it answered just 21.1% of questions correctly and refused to respond to 64% of questions outright. In the more advanced engineering programming test FrontierCode 2.0, its score is 27.2%, lagging behind Haiku 5.5’s 34.9%. Ember is now live on Muse Code and the Meta Model API. Its standard pricing is $0.10 per million input tokens and $0.50 per million output tokens; per-unit prices triple for requests with over 200,000 input tokens.