Meta ships Muse Ember, $0.10-per-million-token AI model

Muse Ember rounds out a three-tier lineup for Meta: Muse Wonder for the hardest problems, Muse Spark for coding and agent work, and Ember for high-volume repetitive tasks. Meta describes pairing the first two as "Wonder plans, Ember does the volume," and notes Ember is the first model trained using Muse Wonder Full, the upgraded version of Wonder.

The model is aimed squarely at work agents repeat all day — calling tools, running tests, making targeted fixes, pulling records and updating systems across apps. Its 1 million token context and native handling of screenshots, graphs, scanned documents, audio and video mean an agent can read a codebase or document set in one call without switching to another model.

Meta frames the release around its stated goal of "personal superintelligence" for everyone, arguing that frontier capability has to be offered at a price that matches scale. Before release, Meta says it evaluated chemical and biological, cybersecurity and loss-of-control risks under its own framework, all below its high-risk thresholds, and checked whether the model resists hidden instructions and asks for human judgment before high-impact actions.