Key Info

Arena reports that Meta's Muse Ember (Max) leads GPT-6 Luna (Max) from OpenAI and Claude Haiku 5.5 (High) from Anthropic across all measured dimensions in its Agent Arena benchmark for efficient models built for real work.

Highlights

  • Speed: 27 seconds per task for Ember vs 5.8 minutes for Luna and 1.9 minutes for Haiku 5.5 — roughly 13x faster than Luna.
  • Performance: +2.3% Net Improvement on the Agent Arena score, vs +1.4% for Luna and +0.4% for Haiku 5.5.
  • Cost: 1.9¢ per task for Ember vs 7.0¢ and 6.9¢ — less than a third of the cost of either competitor.
  • Arena also shares analysis of Ember's steerability score across Work and a look at real-world tasks users are running with Muse Ember in the Agent Arena.