Meta releases Muse Ember for agent execution

Meta describes Ember as its smallest and most efficient model, purpose-built for subagent work such as coding and document analysis.[1] Its listed tasks include tool calls, coding subtasks, routing, extraction and long-document parsing.[1] Meta says Ember can carry a business workflow across apps by retrieving records, calling tools in sequence and updating systems.[1] The company recommends pairing it with Muse Wonder: Wonder handles planning, while Ember takes on high-volume work.[1] That division gives teams a specific way to think about assigning work, but Meta’s description does not establish that every workflow can be automated end to end.[1]

Meta says Ember natively understands text, images, audio and video, so an agent can work with those media without a separate model.[1] The listed context window is one million tokens, with outputs up to 512,000 tokens.[1] Meta makes Ember available on Muse Code and the Meta Model API under the identifier muse-ember-1.0, and says it is also available on Google Cloud and Oracle Cloud.[1] These details give teams several deployment routes and a stated capacity for long inputs; they do not establish availability in other services or support for formats beyond those Meta lists.[1]

Meta lists standard prices of $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens.[1] For prompts longer than 200,000 tokens, the listed rates are $0.30 per million input tokens, $0.012 per million cached input tokens and $1.50 per million output tokens.[1] The higher long-prompt rates matter when estimating costs for document-heavy agent workloads; the standard input price should not be applied to those longer prompts.[1] Separately, Artificial Analysis lists Ember (max) at $0.14 per task and says it was the cheapest model scoring 36 or higher in its comparison.[2]

Artificial Analysis reports that Ember scored 46 on its Coding Agent Index in Muse Code, ahead of Haiku in Claude Code at 37 and Luna in Codex at 31.[2] But on FrontierCode 2.0, Ember scored 27.2%, below Claude Haiku 5.5’s 34.9%.[2] Meta says its safety evaluations found Ember below high-risk thresholds across the categories assessed, and that it received the same core safety training and testing.[1] These findings describe particular evaluations and Meta’s safety assessment; they do not establish that Ember outperforms competitors on every task or is risk-free in a specific deployment.[2][1]