Meta introduced Muse Ember on Nov 9, 2026, the smallest and cheapest model in its Muse family (Meta, The Next Web). Meta pitches it for repeated, bounded work such as tool calls, coding subtasks, routing and document extraction, with a larger model (Muse Wonder or Muse Spark) doing the planning (Meta).
- per million input tokens
- $0.10
- per million output tokens
- $0.50
- on the Artificial Analysis Intelligence Index
- 36
What happened
Muse Ember is available in Muse Code, Meta’s coding agent, on all its subscription plans, and on the Meta Model API as muse-ember-1.0 (Meta). The prices above cover prompts up to 200k input tokens; longer prompts cost $0.30 and $1.50. It reads text, images, audio and video. Meta says an app built on the OpenAI SDK only needs its base URL changed to call it (Meta).
Meta’s own table does not put Ember first everywhere. On FrontierCode 2.0, which judges whether code changes would be merged, Claude Haiku 5.5 scores 34.9 to Ember’s 27.2 (Meta).
The planner and the workers
Meta’s example splits one job between models. Muse Wonder planned a port of 16 language lexers in Pygments, a Python syntax highlighter, to Rust, and 16 Muse Ember agents each ported one lexer in its own copy of the repo. Meta says all 16 matched the Python output on 383 test inputs and on 62 real files held back until the code was frozen, and the run cost about $3.96, of which $1.84 was Ember (Meta). Meta adds that results vary from run to run.
Every benchmark and cost figure here comes from Meta’s post, including the Artificial Analysis score it quotes.




