FULL STORY

Meta Launches Muse Ember: Small Model, Big Value

Meta released Muse Ember, its smallest and most efficient model, on Nov 10, touting intelligence-per-dollar for always-on agents. Artificial Analysis benchmarks soon showed it matching Haiku 5.5 at about a third of the cost.

2026-11-10 ~ 2026-11-10 · 2 episodes · 34 posts

Episode 1 · Meta Launches Muse Ember, Its Smallest Model Built for Agentic Value (2026-11-10, 17 posts)

On November 10, Meta officially released Muse Ember, its smallest and most efficient model to date, positioned as an efficiency-focused model for round-the-clock repetitive agent work. Its core selling point is "intelligence per dollar"—Meta claims that at any inference tier, its intelligence output per unit cost exceeds that of Claude Haiku 5.5, Luna, and various flash models. Scale AI founder Alexandr Wang also announced the release the same day, stating that Muse Ember costs only one-third as much as comparable models.

Confirmed

  • The release was confirmed via Meta's official blog and the @AIatMeta account, with Zuckerberg also taking part in the announcement
  • Model positioning: everyday agent tasks such as tool calling, coding subtasks, routing, information extraction, and long-document parsing; it can carry workflows across applications, call tools in sequence, run tests, and make targeted code fixes
  • Features a 1 million token context
  • Official figures: outperforms Claude Haiku 5.5 on key agent and coding benchmarks, at a cost of $0.116 per task—roughly one-third of Haiku's cost
  • Available through Meta's Model open channel (per reposted posts)

Not yet confirmed

  • The comparison methodology and specific benchmark details behind the claim of "exceeding Haiku, Luna, and all flash-class models": so far only official statements and secondhand accounts exist, with no independent evaluations
  • The exact division of labor with Scale AI: Alexandr Wang participated in the announcement, but the post does not specify Scale AI's role in the model

Why it matters

  • The cost bottleneck for running agents at scale lies in inference overhead from high-frequency small tasks. Muse Ember targets this scenario at roughly one-third the cost—if the numbers hold, it will directly threaten the market share of Haiku and flash-class models
  • Combining a 1 million token context with tool-calling capability lets even a small model handle complex pipeline work like long-document parsing, marking a further split toward the "small model + specialization" route

Episode 2 · Artificial Analysis Reviews Meta's Muse Ember: Matches Haiku 5.5 at One-Third the Cost (2026-11-10, 17 posts)

Artificial Analysis published its full evaluation of Meta's latest efficiency-tier model Muse Ember on November 10. The core finding: the model scores 36 on the Intelligence Index, tying with Claude Haiku 5.5 (max), while costing roughly one-third per task ($0.14 vs $0.41), making its value proposition stand out.

Confirmed

  • Intelligence Index of 36, on par with Claude Haiku 5.5 (max), at about 1/3 the per-task cost ($0.14 vs $0.41)
  • In Artificial Analysis's own coding agent Muse Code, Muse Ember (max) scores 46 on the Coding Agent Index, beating Haiku 5.5 in Claude Code
  • Its standout agent result is 68.1% on AutomationBench-AA, on par with the in-house flagship Muse Wonder and GPT-6 Astra
  • Agent coding performance is polarized: results vary widely across coding benchmarks
  • Cost structure: input/output token pricing matches Claude Haiku 5.5 (max) and GPT-6 Luna (max), but cached input is 60% cheaper and output tokens cost just one-third of Haiku 5.5's; output tokens dominate costs in agent evaluations, which further amplifies the cost advantage

Why it matters

At a fraction of the real-world cost of comparable models, Muse Ember reaches peer-level or even flagship performance on general intelligence and automation agent tasks. If its coding weaknesses are acceptable, it is highly attractive for high-throughput agent workloads. The review also shows its coding ability is inconsistent, so model selection should be weighed against specific benchmarks.