Meteor Ember: a 1M-context model by Meta that matches Claude Haiku 5.5 at a third of the cost

Meta Superintelligence Labs has launched Meteor Ember, the smallest and most efficient model in its family. This hosted multimodal reasoning model is built specifically for repetitive agent tasks, such as tool calls, coding subtasks, and parsing long documents. It natively accepts text, images, audio, and video, enabling agents to read screenshots and execute desktop apps from images. The model offers a 1M token context window and up to 512K output tokens, which is a significant advantage over its competitors. On the Artificial Analysis Intelligence Index, Meteor Ember scores 36 at maximum effort, tying Claude Haiku 5.5. Despite having identical base token pricing, Ember's cost per task is approximately $0.14, which is about a third of what Claude Haiku 5.5 ($0.41) costs. This efficiency makes it highly appealing for high-volume enterprise workflows. In the agent sector, Ember leads with a score of 37.8 on the AutomationBench-AA benchmark. However, it faces challenges in complex coding tasks, where it scores 27.2 on FrontierCode 2.0-AA, trailing behind Claude. Being a hosted model with closed weights, it is accessible via Meta's Model API and Muse Code, focusing specifically on cloud-based automation.