Meta releases Muse Ember, its cheapest Muse model, built for agent execution at 10 cents per million input tokens

Meta released Muse Ember on Monday, the lightest and most cost-efficient model from Meta Superintelligence Labs, TechCrunch reports. It is built for the work agents repeat all day: tool calls, routing, extraction, coding subtasks and long-document parsing.

Ember costs $0.10 per million input tokens, against $1.25 for the mid-weight Muse Spark, Runtime Wire notes. Meta's case rests on cost per task: it says Ember scores 36 on the Artificial Analysis Intelligence Index, similar to Anthropic's Claude Haiku 5.5 at roughly a third of the estimated cost, and above every reasoning setting of OpenAI's GPT-6 Luna, AlphaSignal reports.

In Meta's own comparisons Ember leads on AutomationBench (37.8%), Terminal-Bench (40.9%) and OSWorld 2.1 (69.5%), but trails Haiku 5.5 on FrontierCode 2.0, 27.2% to 34.9%. These are Meta's figures and have not been reproduced independently.

Ember has a 1 million token context window and takes text, images and audio, so agents can read screenshots, scanned documents or meeting recordings without a separate model. Meta suggests pairing it with Muse Wonder, which plans while Ember executes; it is the first Meta model trained with Wonder as its teacher. Ember is the fourth model under the Muse brand, after Spark, the open-weight Glimmer and Wonder.