Meta has just unveiled Muse Ember, a model presented as ‘the smallest and most efficient for executing complex reasoning’ in the context of repetitive tasks assigned to agents. These agents are artificial intelligence programs capable of chaining multiple actions autonomously. Examples cited by Meta include tool calls, code subtasks, routing, data extraction, and analysis of long documents.
Muse Ember accepts text, images, audio, and video as input. An agent can thus read a screenshot or a scanned document, or process a voice memo and a meeting recording, without needing a second model. Its context window, meaning the amount of information it can process at once, reaches 1 million tokens, or word fragments, with up to 512,000 tokens in output. According to Meta, a complete source code or a set of documents can thus fit into a single call.
For the most difficult tasks, Meta recommends pairing Ember with Muse Wonder, its flagship model. The group summarizes the role distribution as follows: ‘Muse Wonder plans, Muse Ember handles the volume.’ Ember is the first model trained from Muse Wonder Full, the most powerful version of the family. It completes a range that now has three tiers: Muse Wonder for the most difficult problems, Muse Spark for code and agentic workflows, and Muse Ember for large-scale repeatable tasks.
Results Compared to Claude Haiku 5.5 and GPT-6 Luna
Meta compares Ember to Claude Haiku 5.5 and GPT-6 Luna on the Artificial Analysis index, reporting the score relative to the cost per task for each reasoning level. At the maximum setting, Ember scores 36 points, like Haiku 5.5, and outperforms Luna, which scores 28. Meta claims it matches Haiku’s best score for about one-third of the cost, which the company presents as the best cost-effectiveness ratio.
On FrontierCode 2.0, a test that evaluates real engineering tasks based on their ability to be merged into a codebase, Claude Haiku 5.5 retains the advantage. It scores 35%, compared to 27% for Ember.
Access and Pricing
Muse Ember is available in Muse Code and via the Meta Model API, under the identifier muse-ember-1.0. It supports the Responses API. Meta also offers it on Google Cloud and Oracle Cloud, like Muse Spark and Muse Wonder.
For prompts, meaning queries sent to the model, of up to 200,000 tokens, Meta charges per million tokens: $0.10 for input, $0.004 for cached input, and $0.50 for output. Beyond 200,000 tokens, the rates increase to $0.30 for input, $0.012 for cached input, and $1.50 for output.
