Meta has a new smallest model, and it is aimed squarely at the busiest part of modern AI. On Monday, November 9, 2026, Meta's Superintelligence Lab launched Muse Ember, the smallest and most cost-efficient model in the Muse family. It is not built to win benchmark crowns. It is built for the work
AI agents repeat all day: the tool calls, routing decisions and document parsing that make up the repetitive 90% of agent work.
What Muse Ember Actually Does
Ember is designed for token-heavy, repetitive work: tool calls, coding subtasks, routing, data extraction and long-document parsing. It is multimodal, reading text, images, audio and video: screenshots, charts, scanned documents, voice memos and meeting recordings, and it can operate a computer's apps from screenshots. It has a 1-million-token context window and can produce up to 512,000 output tokens, so it can chew through long documents and keep working through long, repetitive jobs without losing the thread.
The Price Is the Story
Ember costs $0.10 per million input tokens and $0.50 per million output tokens: about $0.14 per task on Artificial Analysis' composite benchmark. Meta says it matches
Claude Haiku 5.5's Artificial Analysis Intelligence Index score of 36 at roughly a third of the cost per task. The honest caveat is that this framing is Meta' own: on the FrontierCode 2.0-AA coding benchmark, Haiku 5.5 still wins, scoring 34.9% against Ember's 27.2%. Even so, for high-volume agent work, the price gap is hard to ignore.
Where It Fits in the Muse Family
Ember is the fourth Muse model. Spark arrived in April, the open-weights
Muse Glimmer followed in August, and the frontier model Wonder launched in October. Meta sums up the split in one line: 'Muse Wonder plans, Muse Ember handles the volume.' Ember is available now on the Meta Model API as muse-ember-1.0 and inside Muse Code, Meta's coding agent, and it is also listed on OpenRouter. Unlike Glimmer, its weights are closed. Early testers include Box, Cline, Oracle, Cloudflare and You.com.
Why It Matters
Agent demos are cheap; running agents all day is not. A model that handles the repetitive bulk of agent work at ten cents per million input tokens changes the economics of always-on agents: the assistants, coders and researchers that are supposed to keep working while we do other things. That is the real shift here. The model race is moving from 'biggest and smartest' to 'cheapest per job done,' and Ember is Meta's bet that the winner will be the model that does the grunt work best, and cheapest.
Comments
Post a Comment