On November 9, 2026, Meta dropped Muse Ember, the smallest and most efficient model in its new Muse family. This isn't a frontier model chasing human-level reasoning; it is a workhorse designed for the repetitive, high-volume tasks that define modern AI agents: tool calls, coding subtasks, routing, and extraction. Meta’s pitch is directβ€”Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna, specifically targeting the economic bottleneck of agent deployments.

The Three-Tier Architecture

Ember completes a strategic division of labor within Meta's lineup: Muse Wonder handles the hardest problems, Muse Spark manages coding and agentic workflows, and Muse Ember processes repeatable tasks at scale. Meta explicitly states, "Muse Wonder plans, Muse Ember handles the volume." Notably, Ember is the first model trained using Muse Wonder Full, leveraging a teacher-student pipeline that moves beyond mere training details to become a core product feature.

Pricing and Performance Benchmarks

The cost differential is stark. Ember is priced at $0.10 per million input tokens and $0.50 per million output tokens, with cached input at just $0.004. In contrast, Muse Spark 1.3 costs $1.25 for input and $4.25 for output. On Artificial Analysis's cost-per-task metric, Ember clocks in at approximately $0.14 per task versus $2.39 for Spark, a 17x gap for only a one-point difference on the Intelligence Index (36 vs 37). Ember matches Haiku 5.5's Intelligence Index score of 36 but crushes it on AutomationBench-AA, scoring 37.8 against Haiku's 13.9.

Limitations and Safety Protocols

Meta transparently lists where Ember trails. On FrontierCode 2.0-AA, which evaluates real engineering tasks on mergeability, Ember scores 27.2, significantly behind Haiku 5.5's 34.9. This confirms Ember is optimized for structured, repeatable steps rather than open-ended, complex engineering. Regarding safety, Meta evaluated Ember under its Superintelligence Scaling Framework, ensuring it remains below high-risk thresholds for chemical, biological, and cybersecurity risks. It also underwent specific testing for user intent respect and resistance to hidden instructions in web content.

Key Takeaways

  • Route by step type: Use Wonder for planning, Spark for hard coding, and Ember for high-volume routine steps to optimize costs.
  • Retry logic becomes viable: At $0.14 per task, verification loops and multi-sample checks are no longer luxuries but standard architecture.
  • Verify the harness: Benchmarks like Terminal-Bench 4.0 use model-specific best-case harnesses, so pilot Ember on your specific workload before migrating.

The Bottom Line

The agent economy is shifting from maximizing single-step intelligence to minimizing the cost of repeated steps. Ember is Meta's bet that the bottom of the stack wins, and at $0.10 per million tokens, it is a bet builders can afford to test immediately.