Using AI agents? Meta's new model will save you big
Meta has launched Muse Ember, a small and efficient language model for AI agents, with a context window of one million tokens and high performance in automation tests. The price: 10 cents per million input tokens. Competition with OpenAI remains complex due to cache discounts.
Meta has launched Muse Ember, its smallest and most efficient language model, designed for running subtasks of autonomous AI agents. The model features a context window of one million tokens, support for text, images, video, and audio, and a score of 37.8 points on the AutomationBench-AA automation test – significantly higher than Claude Haiku 5.5 (13.9) and OpenAI's GPT-6 Luna. The price is 10 cents per million input tokens and half a dollar per million output tokens. However, OpenAI has an advantage with an aggressive cache discount of about 90% on repeated tokens, which could make Luna cheaper for repetitive tasks. The launch completes Meta's AI model series, which includes Wonder for complex tasks and Spark for code development.