Meta announced Muse Ember on Nov. 9, 2026, targeting repetitive AI agent work. In a launch post, finkd described a model built for tool calls, coding subtasks, routing, extraction and long-document parsing, and announced availability that day through the Meta Model API and Muse Code.
Artificial Analysis reported a specific cost advantage: At max effort, Ember scored 36 on its Intelligence Index, matching Claude Haiku 5.5 at max effort. The evaluator measured Ember’s cost at $0.14 per task, compared with Haiku’s $0.41, or roughly three times as much.
Lower task costs, not lower token rates
The cost-per-task comparison is distinct from the listed token prices. Artificial Analysis reported that Ember had the same input and output token prices as Claude Haiku 5.5 and GPT-6 Luna. It listed Ember’s rates as $0.10 for input, $0.004 for cached input and $0.50 for output per million tokens.
Artificial Analysis reported that Ember at xhigh effort scored 34 at $0.12 per task, rather than the max-effort configuration’s score of 36 at $0.14.
Coding gains came with trade-offs
Artificial Analysis’s coding results were mixed. At max effort, Ember beat Haiku 5.5 on Terminal-Bench 4.0, scoring 40.9% against 32.8%. It trailed Haiku on FrontierCode 2.0, scoring 27.2% against 34.9%.
Knowledge was another limitation in the evaluator’s results. On AA-Omniscience, Artificial Analysis reported Ember’s accuracy at 21.1%, compared with 34.7% for Haiku 5.5 at max effort and 50.5% for Gemini 3.8 Flash at high effort.
Artificial Analysis attributed Ember’s 21.1% AA-Omniscience accuracy mostly to its declining to answer 64% of questions. It also reported a hallucination rate of 17.4% on that benchmark.
