Meta Releases Cheap Model for AI Agents

Meta's Superintelligence Lab, the company's AI research division, released a small model called Muse Ember on Monday. It is built for the repetitive work AI agents do all day: calling other software tools, parsing code, and reading long documents. Ember also handles images and sound, so an agent can read a screenshot or listen to a voice memo without switching to a separate model. Meta prices it at 10 cents per million tokens, the units AI models use to process text. That is notably cheaper than Anthropic's Haiku and OpenAI's Luna, two rival small models built for similar jobs.

Ember is designed to work alongside Meta's larger Muse Wonder model. Wonder draws up the plan, and Ember carries out the steps in bulk at a lower cost. For companies already making thousands of model calls a day, that price gap can add up to real savings. The release puts Meta in direct competition with Anthropic and OpenAI, not for the best chatbot, but for the job of cheap, high-volume executor inside other companies' AI agent systems.