Meta has unveiled Muse Ember, its smallest model yet, designed to handle repetitive coding and reasoning tasks performed by AI agents while keeping operating costs low.The compact model is built for everyday agent workloads, including tool calls, coding subtasks, information extraction, request routing, and parsing lengthy documents. Meta says Muse Ember delivers more intelligence per dollar than Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna across reasoning settings.According to Artificial Analysis, Muse Ember scores 36 on its Intelligence Index, matching Claude Haiku 5.5 while costing approximately $0.14 per task. This is nearly three times cheaper than Haiku’s estimated $0.41 cost at the same score.The model also compares favorably with Meta’s larger Muse Spark 1.3. Artificial Analysis estimates that Muse Ember costs $0.14 per Intelligence Index task, compared with $2.39 for Spark, which scores 37 on the index.Built for everyday agent tasksMeta designed Muse Ember to handle the numerous smaller operations that agents perform during extended workflows rather than focusing exclusively on complex reasoning problems.It can execute business workflows from start to finish by retrieving records, making sequential tool calls, and updating information across applications. The model also comes with a 1-million-token context window and can generate up to 512,000 output tokens, allowing it to process entire codebases or large document collections in a single call.In evaluations, Muse Ember scored 37.8 percent on AutomationBench-AA, which measures end-to-end business workflows. This compared with 20.2 percent for GPT-6 Luna and 13.9 percent for Claude Haiku 5.5.It also achieved 61.6 percent on Agents’ Last Exam, narrowly exceeding Luna’s 57.1 percent and Haiku’s 57 percent. On the Personal Superintelligence Index, Meta’s internal measure of personal and knowledge work, it scored 46.8 percent, ahead of Luna’s 37 percent and Haiku’s 36.1 percent.Its coding performance was similarly competitive. Muse Ember scored 40.9% on Terminal-Bench 4.0, compared with 39.2 percent for Haiku and 16.4 percent for Luna. However, Haiku performed better on FrontierCode 2.0-AA, scoring 34.9 percent against Muse Ember’s 27.2 percent.Multimodal capabilities and availabilityBeyond coding and agent execution, Muse Ember can process text, images, audio and video. This allows it to interpret screenshots, charts and scanned documents, as well as voice recordings and video content.On OSWorld 2.1, which measures computer-use capabilities across desktop applications, Muse Ember scored 69.5% using partial-credit scoring. It also achieved 80.6 percent on PerceptionBench, which evaluates fine-grained visual understanding with tools.Meta said the model underwent safety evaluations covering chemical and biological risks, cybersecurity, and loss-of-control scenarios. The company reported that Muse Ember remained below its framework’s high-risk thresholds.The model is available through Muse Code and the Meta Model API as muse-ember-1.0. Pricing starts at $0.10 per million input tokens and $0.50 per million output tokens, while cached input costs $0.004 per million tokens.Muse Ember is the first model Meta says was trained using Muse Wonder Full. It joins Muse Wonder, designed for the hardest reasoning problems, and Muse Spark, aimed at coding and agentic workflows.Meta’s broader goal is to make advanced capabilities affordable enough to run across everyday products and services, with Muse Ember targeting the repetitive workloads that agents handle at scale.