Meta launches Muse Ember, a small model built for AI agent work at $0.10 per million input tokens

Meta on Monday (November 9) unveiled Muse Ember, the smallest and most efficient model in its Muse family, designed for the work AI agents repeat all day: tool calls, coding subtasks, request routing, extraction and long-document parsing. It is available now on the Meta Model API and in Muse Code under the name muse-ember-1.0, and is also listed on OpenRouter. Meta says Ember performs competitively against models from Anthropic and OpenAI in agent workflows and coding. (Image: file image of a Meta data center.)

Built for agents' repetitive work

Unlike flagship models that chase maximum reasoning power, Ember's pitch is cost per finished task. It accepts text, images, audio, video and PDF inputs, returns text, and offers a 1-million-token context window with outputs of up to 512,000 tokens, enough to take in a full codebase or document set in one call. Meta also says Ember is the first model trained using Muse Wonder Full, and that Ember, Spark and Wonder are now available on Google Cloud and Oracle Cloud.

Pricing and specifications

Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. On the Artificial Analysis benchmark that works out to about $0.14 per task at its highest reasoning setting. The weights are private: the model runs only through Meta's API and services that resell it, and cannot be downloaded.

How it compares

On the Artificial Analysis Intelligence Index at maximum reasoning, Ember scores 36, close to Anthropic's Claude Haiku 5.5 at 36.2, but Haiku costs about $0.41 per task, roughly three times as much; OpenAI's GPT-6 Luna costs about $0.15 per task. Against Meta's own larger Muse Spark 1.3, Ember scores 36 versus 37, while Spark costs $2.39 per task. Meta says Ember delivers more intelligence per dollar than Haiku 5.5 or GPT-6 Luna at every reasoning setting.

Anthropic launched on the same day

Notably, Anthropic unveiled Fable 5.2 and Mythos 5.2 on the same day, aimed at more demanding technical work, and cut Fable's price to $8 per million input tokens from $10. The two launches on one day show the AI race shifting from who is smartest to who can do sustained, complex work at the lowest cost.