Meta launches Muse Ember to compete with Claude Haiku and GPT-6 Luna

A new tier of AI efficiency

Meta recently introduced Muse Ember, a smaller and more efficient member of its Muse model family. This release follows a broader strategy by the company to offer a variety of models tailored for different needs. While larger models like Muse Wonder Full handle complex planning, Ember is designed for high-volume, focused tasks.

The company is positioning this new model to compete directly with other small-scale offerings from major labs. Specifically, it is intended to rival OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5. By creating a specialized tier for smaller tasks, Meta aims to provide a more versatile toolkit for developers building AI applications.

Ember is currently available through the Meta Model API under the ID muse-ember-1.0. It is also accessible via Muse Code. The company has indicated that it may soon appear on other platforms like OpenRouter and various neoclouds, following the pattern of its previous Muse releases.

Benchmarking and cost analysis

One of the primary goals for Muse Ember is to provide high performance at a lower price point than its competitors. According to the Artificial Analysis Intelligence Index, Ember costs about $0.14 per task at its max reasoning setting. This is significantly cheaper than the $0.41 per task cost for Claude Haiku and slightly less than the $0.15 per task cost for GPT-6 Luna.

The pricing structure for Ember is designed to be competitive with standard short-context rates. It costs $0.10 per million input tokens and $0.50 per million output tokens. These rates match the standard prices for both GPT-6 Luna and Claude Haiku 5.5, while offering a lower cost for cached input tokens compared to those rivals.

When looking at Chinese competitors, Ember undercuts the uncached input rates for models like DeepSeek V4.1 Flash 1017 and Z.ai’s GLM-5.3 Flash. It also matches the output prices of GLM while remaining cheaper than DeepSeek's off-peak rates. These comparisons suggest Meta is aggressively targeting the budget-conscious segment of the AI market.

Multimodal capabilities and training

Muse Ember is a multimodal model, meaning it can process various types of data beyond just text. It supports inputs for text, images, audio, and video. One specific capability Meta highlighted is the model's ability to operate desktop applications by interpreting screenshots provided as visual input.

The development of Ember involved a specific training methodology used by Meta to improve its smaller models. The company reported that Ember was the first model it trained using the larger Muse Wonder Full model. This technique is similar to how the company developed Muse Wonder to power its personal agent and Meta AI services.

In practical testing on the OSWorld 2.1 benchmark, Ember demonstrated its ability to handle computer use tasks. It earned a partial-credit score of 69.5 percent, which is nearly identical to the 68.9 percent scored by Claude Haiku. However, Ember finished more tasks outright, completing 33.8 percent of tasks compared to Haiku's 27.8 percent.

Performance trade-offs

While Ember excels in many areas, it does not lead every single benchmark. For example, on the FrontierCode 2.0 benchmark, which measures whether a repository maintainer would merge an agent's changes, Ember scored 27.2 percent. This was lower than the 34.9 percent achieved by Claude Haiku and the 28.5 percent scored by DeepSeek V4.1 Flash 1017.

Meta also tracks its own internal metrics, such as the Personal Superintelligence Index. On this specific measure of personal and knowledge work, Meta reports that Ember holds a lead over both Haiku and Luna. These internal scores help the company evaluate how well the model performs on specific types of human-centric tasks.

The model's performance also varies depending on the settings used. At the xhigh setting, Ember scores about 34 on the Intelligence Index at a cost of roughly $0.12 per task. These variations allow developers to choose the specific balance of reasoning power and cost that fits their specific application needs.