Meta releases Muse Ember, its lightest and cheapest model, built for agent sub-tasks
Priced at 10 cents per million input tokens, the multimodal model is meant to handle tool calls, coding subtasks, routing and document parsing while Muse Wonder plans.
Why it matters
Ember is the fourth model under Meta's Muse brand, after Spark in April, the open-source Glimmer in August and the frontier Wonder model in October. Meta pitches it on cost per task, citing Artificial Analysis scores that put it ahead of Anthropic's Haiku and OpenAI's Luna, and says it reads text, images and audio without a separate model. The split, where a large model plans and a small one does the volume, is how many agent products are now built. Watch independent cost-per-task results and whether Meta releases open weights.
Key facts
- Released Nov 9, 2026 by Meta Superintelligence Lab
- $0.10 per million input tokens
- Fourth Muse model after Spark, Glimmer and Wonder
- Handles text, images and audio inputs
- Designed to pair with Muse Wonder in agent harnesses