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

Beating AI News Flash: Meta has released Muse Ember, its lightest AI model to date, specifically designed to handle repetitive agent tasks such as tool calls, code modifications, and document analysis. Ember supports text, image, audio, and video inputs, with a context window of 1 million tokens.

On Artificial Analysis's Intelligence Index, Muse Ember scored 36, tying with Claude Haiku 5.5, but at a cost of only $0.14 per task—roughly one-third of the latter's. Compared to Meta's own Muse Spark 1.3, Ember scores just 1 point lower while costing approximately 1/17 as much.

The key to its low cost lies in token utilization efficiency. Ember and Haiku 5.5 actually have the same standard input and output pricing, but Ember's cache read price is 60% lower, and it generates only about one-third of Haiku's output tokens when completing evaluation tasks.

Meta plans to have the large model Muse Wonder handle planning for complex tasks, then delegate large volumes of repetitive work to Ember. For example, Wonder could devise a code migration plan, while multiple Ember agents modify files and run tests in parallel.

However, Ember has notable weaknesses. In Artificial Analysis's knowledge test, it answered only 21.1% of questions correctly, while refusing to answer 64% outright. On the more complex engineering coding test FrontierCode 2.0, it scored 27.2%, trailing Haiku 5.5's 34.9%.

Ember is now available on Muse Code and the Meta Model API. Standard pricing is $0.10 per million input tokens and $0.50 per million output tokens, but for single requests exceeding 200,000 input tokens, the unit price triples.