Muse Ember, AgentBaton and CPU-ARC: put the cheap model where the volume is

Date: 2026-11-10 · Sources: Meta Muse Ember blog (via @AIatMeta, @shengjia_zhao, The Decoder) · AgentBaton (arXiv 2610.35959) · CPU-ARC (arXiv 2611.09891) · Inkling 2 Small (Arena) · raw: raw/twitter/feed/2026-11-10-morning-ranked.json, raw/huggingface/2026-11-09-agentbaton-*.md, raw/huggingface/2026-11-09-cpu-arc-*.md

TL;DR

• Muse Ember is Meta's smallest Muse model, built for what agents repeat all day: tool calls, coding subtasks, routing, extraction, long-document parsing. Price: $0.10 per million input tokens, $0.50 per million output. On the Artificial Analysis Intelligence Index it scores 33.9 at $0.116 per task and 35.8 at $0.136. Claude Haiku 5.5 needs $0.248 per task to reach 34.3, so Ember matches Haiku at about a third of the cost, and beats every GPT-6 Luna setting. Agents benchmarks (max effort): AutomationBench-AA 37.8 vs Haiku 13.9, OSWorld 2.1 69.5 partial, Terminal-Bench 4.0 40.9 vs 39.2. It is weaker on FrontierCode (27.2 vs Haiku's 34.9). Meta pitches pairing it with the larger Muse Wonder: "wonder can plan, ember can handle the volume." It ran as the anonymous "crouton-alpha" and processed over 5T tokens in testing.
• AgentBaton asks whether two models can share one long agent task. Handing off from cheap to capable mostly fails: little saving, lower accuracy. Handing off from capable to cheap at the right moment keeps or improves accuracy. A small learned detector picks that moment from task-progress patterns. Across 11 model pairs and two benchmarks it cuts API cost by 65.9% on average.
• CPU-ARC (agentic reasoning caches) separates solving a task from applying the solution. Once an LLM has reasoned through a task type, the trace is compiled into a CPU-executable cache that handles new instances. It matches frontier LLMs on hard math and science tasks at 16-291x lower invocation cost.

flowchart LR
  T["Long task<br/><small>many steps</small>"] --> B["Capable model<br/><small>plans, early steps</small>"]
  B --> D{"Handoff detector<br/><small>AgentBaton</small>"}
  D -->|stable phase| E["Cheap model<br/><small>Ember-class</small>"]
  D -->|still hard| B
  E --> C["Reasoning cache<br/><small>CPU-ARC replay</small>"]
  C --> O["Result<br/><small>65-99% cheaper</small>"]
  E --> O
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class T input
  class B,E core
  class D loop
  class C,O exit
  linkStyle 2 stroke:#2f9e44,stroke-width:2px
  linkStyle 3 stroke:#f08c00,stroke-width:2px

Relation to prior wiki pages

• Pattern (N-of-a-kind, now four): lead-plus-workers routing keeps showing up. Cohere's Nouha Dziri talk (11-09: frontier lead plus cheaper workers nears an all-frontier coding team at about 70% lower cost), HKU's AgentPilot (11-07: a 9B model chooses strategy and stopping while a frontier model executes), Claude Code routing status updates to Haiku (11-09: about 99% cheaper per update), and now AgentBaton's capable-to-cheap handoff (65.9% savings). The open question is no longer "should we route inside a task" but "when."
• Contradiction: AgentPilot puts the small model in charge of control decisions; AgentBaton finds small-first handoffs fail. They differ on what the small model does (control vs execution), which may reconcile them. Unresolved.
• CPU-ARC is the strongest version of the decision-reuse idea from HarnessMM (10-27: the harness skips LLM calls whose outcome it already planned, 88% fewer tokens). CPU-ARC removes the LLM from repeated task types entirely.
• Inkling 2 Small (Thinking Machines, weights coming): #31 in Agent Arena, would be #9 among open models, $0.25 median per task, 70% below Kimi K3. A second cheap-volume model in the same window.
• Concept: llm-routing.

Gaps

Ember's numbers are Meta's, at "proposed pricing" estimates on the AA chart. AgentBaton's detector is trained per pair; transfer to unseen pairs is not shown in the abstract. CPU-ARC's "robust to perturbations" claim needs checking on tasks whose structure varies, not just their numbers.