Timeline

Meta releases Muse Ember, a small model for agents

Priced at $0.10 and $0.50 per million tokens; Artificial Analysis scored it 36 on its index at $0.14 per task, level with Claude Haiku 5.5 at maximum effort for about a third of the cost.

  • Models & capabilities
  • Minor

Meta released Muse Ember, which it called its smallest and most efficient model, aimed at the repeated work of agents: tool calls, coding subtasks, routing, extraction and long-document parsing. It is the first model Meta trained using Muse Wonder Full, the larger system it released in October. Meta’s post says Ember has a 1-million-token context window and up to 512,000 tokens of output, and costs $0.10 per million input tokens and $0.50 per million output tokens.

The weights are not open. Meta’s post describes access only through the Meta Model API and Muse Code, plus Google Cloud and Oracle Cloud listings, and offers no download. That sets it apart from the open-weight Muse Spark variant Meta released in October, and from the open models TechCrunch noted it competes with on price.

Meta’s headline claim is that Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting. Artificial Analysis, the benchmarking firm whose data Meta used, published its own scoring the same day: Ember at maximum effort scored 36 on its Intelligence Index at $0.14 per task, the same score as Haiku 5.5 at maximum effort at $0.41. Meta’s own chart gives 35.8 against 36.2, and notes that Ember’s costs are estimates at proposed pricing. Artificial Analysis put Muse Wonder at 44 for $6.95 per task.

The independent write-up is mixed beyond the headline. Ember beat Haiku 5.5 on Terminal-Bench 4.0 (40.9 per cent against 32.8 per cent) but trailed it on FrontierCode 2.0 (27.2 against 34.9). On Artificial Analysis’s knowledge test, Ember’s accuracy of 21.1 per cent was the lowest among the similarly capable models it compared, largely because it declined to answer 64 per cent of questions; its hallucination rate, at 17.4 per cent, was about half Haiku’s. Some of the other results in Meta’s table are Meta’s internal evaluations rather than third-party runs.

Meta said it assessed Ember under its Meta Superintelligence Scaling Framework and found it below the framework’s high-risk thresholds, and that more efficient models are in training.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.