Muse Ember (Max) is a proprietary reasoning model developed by Meta, released in November 2026. According to the Artificial Analysis Intelligence Index, it achieves a score of 36, significantly outperforming the median score of 24 for comparable models.

Core Methodology and Evaluation

The model’s intelligence is quantified via the Artificial Analysis Intelligence Index (v5.0.1), a composite benchmark that aggregates performance across nine distinct domains: AA-Briefcase v1.1 (agentic knowledge work), AutomationBench-AA (SaaS workflows), Terminal-Bench 4.0 (coding/terminal use), FrontierCode 2.0 (repository-level software engineering), AA-Omniscience v1.1 (knowledge reliability/hallucination rates), GDP.pdf (document reasoning), HLE-Diamond (reasoning and knowledge), CritPt v2 (physics reasoning), and Terminal-Bench-Science 0.1 (scientific research workflows).

The methodology utilizes a weighted average cost and performance framework. The total weighted cost per task ($C_{task}$) is derived from the costs associated with input, cache hits, cache writes, reasoning, and output tokens ($T_i, T_c, T_w, T_r, T_o$):

$$C_{task} = \sum_{j=1}^{9} w_j \left( \frac{Input_j \cdot P_i + Cache_j \cdot P_c + Output_j \cdot P_o}{N_{tasks}} \right)$$

where $w_j$ represents the weight of evaluation $j$. The performance is measured via "Time per Intelligence Index Task," calculated as the ratio of output tokens produced per task to the model's end-to-end output speed.

Performance and Technical Specifications

Muse Ember (Max) is categorized as a "reasoning" model, meaning it employs extended chain-of-thought processing before yielding an output. Key technical benchmarks include:

• Latency: A time-to-first-answer-token of 4.37s and an output speed of 446 tok/s.
• Efficiency: It completes tasks in 3.2 minutes, which is faster than the median of 7.5 minutes for its peer group.
• Verbosity: The model generates 115M output tokens across the index, indicating a higher-than-average verbosity (median: 90M).
• Context: It supports a 1M token context window, equivalent to approximately 1,573 pages of 12-point Arial font.
• Modality: The model is multimodal, accepting both text and image inputs while providing text outputs.

Economic Positioning

Muse Ember (Max) is positioned as a competitively priced, high-intelligence model. Its pricing structure is $0.10 per 1M input tokens and$0.50 per 1M output tokens, with a 96% cache discount. Compared to market medians of $0.24 and$0.80 respectively, it serves as a high-performance option for enterprise-grade reasoning and agentic workflows.