Introducing Our Smallest, Most Efficient Model: Muse Ember

Open original

Meta is releasing Muse Ember, a small, low-cost model aimed at the repetitive work agents do constantly, like tool calls, coding subtasks, and document extraction. The company claims it beats comparable Claude and GPT models on cost-per-task efficiency, with benchmark numbers that are mostly Meta's own or from Artificial Analysis.

  • Priced at $0.10 per million input tokens and $0.50 per million output, with a 1M-token context window, and Meta puts its cost at roughly $0.14 per Artificial Analysis task.
  • On the benchmarks Meta highlights, Ember leads Claude Haiku 5.5 and GPT-6 Luna on agentic tasks like AutomationBench (37.8% vs 13.9% and 20.2%) and Terminal-Bench (40.9% vs 39.2% and 16.4%), but it trails Haiku on FrontierCode 2.0 (27.2% vs 34.9%).
  • Several headline comparisons come from Meta's internal evals, such as the Personal Superintelligence Index and OSWorld 2.1, so they are harder to verify independently than the Artificial Analysis numbers.

Meta is launching Muse Ember, a small, cheap model aimed at the grunt work agents do constantly: tool calls, coding subtasks, routing, extraction, and parsing long documents. Meta's headline claim is about cost efficiency. Per Artificial Analysis data dated November 9, 2026, Ember matches Claude Haiku 5.5's top score (36) at roughly a third of the cost, and it beats GPT-6 Luna at every reasoning setting. Its cost per task runs from $0.116 to $0.136 across the settings Meta shows, against $0.092 to $0.411 for Haiku and $0.073 to $0.152 for Luna. API pricing is $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens. Note that these per-task figures are Meta's estimates at proposed pricing, and the comparison is drawn from a single benchmark aggregator. On the task-level evals, Ember's wins are uneven. It leads on AutomationBench-AA business workflows (37.8% against 20.2% for Luna and 13.9% for Haiku), on Agents' Last Exam (61.6%), and on Meta's internal Personal Superintelligence Index (46.8%). It also tops Terminal-Bench 4.0 (40.9%, just ahead of Haiku's 39.2%) and the Coding Agent Index (46 versus 37 and 31). It does not win everywhere. On FrontierCode 2.0, which grades real engineering tasks on mergeability, Haiku scores 34.9% and Ember 27.2%. The multimodal results are close: Ember and Luna tie at 22.8% on GDP.pdf, and Ember edges Haiku on OSWorld 2.1 (69.5% partial credit against 68.9%, but 33.8% against 27.8% on binary scoring). Several of the benchmarks are Meta-internal or partially scored, so the "best result" labels deserve some skepticism until independent replication. Meta also pitches Ember as a pairing piece rather than a standalone. Its suggested setup is Muse Wonder for planning and Ember for the high-volume execution, which is a sensible architecture if the cost claims hold up. The practical question for anyone building agents is whether Ember's advantage survives contact with their own workloads, especially the harder engineering tasks where it trails Haiku. The post cites a fuller report for methodology, so that's the document to read before making decisions on the numbers.

Ask about this article