Meta AI has released Muse Ember, a small, efficient model targeting the same agentic workload segment as GPT-6 Luna and Claude Haiku 5.5—specifically tool calling, coding subtasks, routing, extraction, and long document parsing. All three models share a base rate of $0.10 per million input tokens and $0.50 per million output tokens for shorter prompts, but diverge beyond their respective context thresholds: Ember’s threshold is 200k tokens (above which input rises to $0.30/M and output to $1.50/M), Luna’s is 272k ($0.20/M input, $0.75/M output), and Haiku 5.5’s is 100k ($0.50/M input, $2.50/M output). Cached input pricing also differs: Ember at $0.004/M (≤200k) and $0.012/M (>200k), Luna at $0.01/$0.02, Haiku at $0.01/$0.05. Muse Ember offers six reasoning levels (minimal, low, medium, high, xhigh, max). In Artificial Analysis benchmarks, Ember sits on the Pareto efficiency frontier, meaning no model achieves a higher Intelligence Index score for a lower cost per task; it matches the performance of the far more expensive Muse Wonder (40× the price) on qualitative tests such as SVG generation, though reasoning level did not consistently correlate with visual output quality. The high reasoning level failed to return a complete SVG in testing.