Goal presented Muse Ember, its most compact and cost-effective artificial intelligence model to date. It focuses on repetitive tasks assigned to AI agents, with an entry price of $0.10 per million tokens.
The model is now available in Muse Code and through the API under the name muse-ember-1.0. It has a context window of one million tokens and can generate up to 524,288 tokens in a single response. This capability allows you to process large code bases or sets of documents without having to split them into multiple queries.
Muse Ember was designed to support operations that an AI agent frequently performs: calling tools, extracting information, organizing tasks, analyzing large documents, or even performing programming subtasks. Meta relies on this specialization to accelerate volume processing without systematically mobilizing a more imposing model.
The model is also multimodal. Accepts text, images, audio and video. This way, an AI agent can use a screenshot, chart, scanned document, or meeting recording without using a separate model for each format. Meta also claims that Ember can interact with a computer from screenshots.
Muse Ember focuses on complementary operation
Meta’s strategy is to pair Muse Ember with a more powerful model like Muse Wonder. The latter can plan an operation and define the rules to follow before delegating its application to Muse Ember in a large number of cases. Meta summarizes this distribution as follows: “Muse Wonder plans, Muse Ember handles the volume”. Ember becomes the fourth model in the Muse family, following Spark, its open source counterpart Glimmer and Wonder.
Its price constitutes one of its main arguments. Meta charges $0.10 per million incoming tokens, $0.50 per million generated tokens, and $0.004 per million cached incoming tokens. According to the company’s calculations, a task evaluated using Artificial Analysis costs about $0.14 with Muse Ember, compared to $2.39 with Muse Spark 1.3.
Encouraging results, but uneven performance
Meta highlights favorable results against Claude Haiku 5.5 and GPT-6 Luna in various agent and programming tests. In the Artificial Analysis Intelligence Index, Muse Ember scores 36 points, tied with Claude Haiku with 5.5 and ahead of GPT-6 Luna which has 28 points. However, these results should be interpreted with caution: some benchmarks are internal to Meta and test environments may vary.
Vals AI’s independent benchmarks provide more information. Muse Ember is ranked 12th out of 35 models in the Vals index with a score of 60.48%. It performs especially well in the legal research tests, where it comes in fourth place out of 36 models. Its results are less convincing in Terminal-Bench 4.0, ranking 15th out of 62 models.
With Muse Ember, Meta aims to make running AI agents more affordable at scale. The model is not intended to replace the most powerful systems in its range: it should allow them to delegate numerous and repetitive operations to a less expensive tool.
