Source · article
Muse Ember
Summary
Muse Ember (muse-ember-1.0) is Meta's smallest and cheapest Muse model, released on
2026-11-09 for agent sub-work: tool calls, coding subtasks, routing, extraction, and
long-document parsing. It is the first model trained with Muse Wonder Full. It is
API-only: Meta Model API (including a Responses-style API), Muse Code, Google Cloud,
and Oracle Cloud. The page does not mention open weights or a licence.
Key points
- 1M-token context, up to 512K output tokens; text, image, audio, and video input; reasoning levels medium, high, xhigh, and max.
- Price per 1M tokens: $0.10 input, $0.004 cached input, $0.50 output for prompts up to 200K; $0.30 / $0.012 / $1.50 above 200K.
- Meta's figures at max reasoning, against GPT-6 Luna and Claude Haiku 5.5: Terminal-Bench 4.0 40.9% (39.2%, 16.4%), OSWorld 2.1 69.5% (68.9%, 50.5%), AutomationBench-AA 37.8% (20.2%, 13.9%); behind Haiku 5.5 on FrontierCode 2.0-AA (27.2% vs 34.9%). Artificial Analysis Intelligence Index 36, against 37 for Muse Spark 1.3 at about 17 times the cost per task.
- Safety: below high-risk thresholds in Meta's scaling framework for the categories tested; tested for following user intent and resisting injected instructions.
Why it matters
A low-cost sub-agent model with a 1M context, priced like Claude Haiku 5.5's lowest tier, gives another option for the cheap worker model in planner/worker setups. The benchmarks come from Meta and have not been independently checked.