- Muse Ember is Meta's smallest AI model, designed for low-cost handling of everyday tasks.
- The model outperforms Claude Haiku 5.5 and GPT-6 Luna in terms of intelligence index and cost.
- Muse Ember can perform multiple small operations and has strong contextual processing capabilities.
- The model supports processing of various media types and has undergone safety assessments, maintaining a high-risk threshold.
Meta has released its smallest model, Muse Ember, designed to handle repetitive coding and reasoning tasks performed by AI agents while maintaining low operational costs. This compact model is suitable for everyday agent workloads, including tool invocation, coding subtasks, information extraction, request routing, and parsing lengthy documents. Meta claims that Muse Ember provides more intelligence per dollar in reasoning settings than Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna.
According to a report by Artificial Analysis, Muse Ember scored 36 on its intelligence index, the same as Claude Haiku 5.5, while the cost per task is approximately $0.14 (about HK$1.09), nearly three times lower than Haiku's estimated cost of $0.41 (about HK$3.20). The model also performed well in comparison to Meta's larger Muse Spark 1.3.
Artificial Analysis estimates that the cost per intelligence index task for Muse Ember is $0.14 (about HK$1.09), while Spark's cost is $2.39 (about HK$18.63), with an intelligence index score of 37.
Versatility of Muse Ember
Meta designed Muse Ember to handle multiple small operations executed by agents in extended workflows, rather than focusing solely on complex reasoning problems. The model can execute business workflows from start to finish by retrieving records, making sequential tool calls, and updating information in applications. It is also equipped with a context window of 1 million tokens and can generate up to 512,000 output tokens, allowing it to process an entire codebase or a large collection of documents in a single call. In evaluations, Muse Ember scored 37.8% on AutomationBench-AA, a metric measuring end-to-end business workflows, while GPT-6
Luna scored 20.2%, and Claude Haiku 5.5 scored 13.9%. Additionally, Muse Ember scored 61.6% on Agents’ Last Exam, slightly higher than Luna's 57.1% and Haiku's 57%. In Meta's personal super intelligence index, it scored 46.8%, ahead of Luna's 37% and Haiku's 36.1%. Its coding performance is also competitive, with Muse Ember scoring 40.9% on Terminal-Bench 4.0, while Haiku scored 39.2%,
Luna scored 16.4%. However, on FrontierCode 2.0-AA, Haiku performed better, scoring 34.9%, while Muse Ember scored 27.2%.
Pricing and Usage
In addition to coding and agent execution, Muse Ember can handle text, images, audio, and video. This enables it to interpret screenshots, charts, and scanned documents, as well as voice recordings and video content. In OSWorld 2.1, the model utilized partial credit scoring, achieving a score of 69.5%. Its score in PerceptionBench was 80.6%, a metric assessing fine-grained visual understanding capabilities. Meta stated that the model has undergone safety assessments covering chemical and biological risks, cybersecurity, and uncontrolled scenarios, with the company reporting that Muse Ember consistently remains below its framework's high-risk threshold.
The model can be accessed through Muse Code and the Meta Model API under the name muse-ember-1.0, with pricing starting at $0.10 (about HK$0.78) per million input tokens and $0.50 (about HK$3.90) per million output tokens, while the cost for cached input is $0.004 (about HK$0.03) per million tokens. Muse Ember is Meta's first model claimed to utilize Muse Wonder Full training.
It was launched alongside Muse Wonder, designed for the most challenging reasoning problems, and Muse Spark, aimed at coding and agent workflows. Meta's broader goal is to make advanced capabilities cost-effective enough to operate in everyday products and services, with Muse Ember targeting repetitive workloads processed at scale by agents.
Impact of Muse Ember in the AI Field
The launch of Muse Ember marks a significant advancement for Meta in AI model development, particularly in reducing costs and enhancing efficiency. This model not only handles coding and reasoning tasks but also supports the interpretation of multimedia such as text, images, and audio, demonstrating its versatility. As AI technology becomes increasingly integrated into daily life, the low-cost features of Muse Ember will help drive broader commercial adoption and may alter corporate strategies in automation and data processing.

