Meta Muse Ember: smallest efficiency-tier agent model on Meta Model API at $0.10/$0.50

Why it matters

【Key points】

• On 2026-11-09 Meta primary-launched Muse Ember: smallest, most efficient Muse model for agent steps that repeat all day—tool calls, coding subtasks, routing, extraction, long-document parsing.
• Research blog + @AIatMeta live same day; Meta Model API at $0.10/$0.50 per M input/output tokens.
• 1M context, up to ~512K output; natively multimodal.
• Meta says it matches Claude Haiku 5.5’s best AA score at ~1/3 cost and beats every GPT-6 Luna setting; AutomationBench-AA (max) self-reported 37.8 vs Haiku 13.9 / Luna 20.2.
• Portfolio: Wonder for hardest problems, Spark for coding/agents, Ember for volume (~$0.14/AA task).
• Partner quotes: Box, Cline, Oracle, Cloudflare, You.com.
• API/product launch, not open weights.
• Site has Wonder/Spark Open/Glimmer; no Ember.

【Our read】

Background

Muse is filling the agent token-economics tier—most agent spend is many small tool steps, not one flagship call. Ember is an efficiency API SKU with 1M context and multimodal inputs.

Peers

• vs Haiku 5.5 / GPT-6 Luna, Meta charts lower $/task with similar or higher Index
• vs Wonder/Spark Open on site, Ember is not a weights event—“Wonder plans, Ember handles volume.” vs Chinese flash/small models, delta is closed API + Meta Superintelligence Labs branding
• wait for third-party harnesses.

Impact

• near-term pressure on subagent/doc-review/high-throughput coding costs
• watch benches，real tool-call bills，open weights vs API-only，Haiku/Luna price response.

【Verifiable】

• BlogdatePublished 2026-11-09T00:00:00.000Z
• @AIatMeta2026-11-09T18:04:15Z
• Pricing
• AA secondarysecondary only