Meta summarizes its dual agent strategy with the motto “Muse Wonder plans and Muse Ember handles the heavy lifting,” introducing a multimodal model that matches Claude Haiku 5.5 benchmarks at roughly one-third of the task cost.
Launch of Muse Ember for repetitive agent tasks and competition with Haiku and Luna
Pricing structure and task completion cost on the Artificial Analysis index
Diverging results in coding and OSWorld computer-use benchmarks
Multimodal integration with Muse Wonder and expansion of the Muse model family
Frequently asked questions
Launch of Muse Ember for repetitive agent tasks and competition with Haiku and Luna
Meta Superintelligence Labs launched the “Muse Ember” model on Monday, November 9, 2026, describing it in the official Meta Research blog as its lightest and most cost-effective AI model to date. According to TechCrunch, the model is designed “to perform complex reasoning and all-day repetitive agent work: tool calls, sub-coding tasks, routing, data extraction, and long-document analysis.”
This launch gives Meta a direct competitor to Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna in the market for high-density, economical models that handle repetitive work within agent ecosystems.
Pricing structure and task completion cost on the Artificial Analysis index
Muse Ember is priced at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, boasting a 1-million-token context window with an output capacity of up to 512,000 output tokens, according to The New Stack. While standard input and output pricing matches what Luna and Haiku charge for short prompts, Meta’s bet relies on the total cost per completed task.
At maximum reasoning effort, Meta estimates Ember’s cost at about $0.14 per task on the Artificial Analysis Intelligence Index, compared to roughly $0.41 for Haiku and $0.15 for Luna. On that index, Ember scores 36 points, matching Haiku and outperforming Luna’s 28 points, while also competing on pricing with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash.
Diverging results in coding and OSWorld computer-use benchmarks
Not all benchmark results favored the new model. In the FrontierCode 2.0 test, which measures whether a repository maintainer would accept an agent’s code modifications, Ember scored 27.2%, trailing Haiku’s 34.9%.
Conversely, the model outperformed competitors in the OSWorld 2.1 computer-use benchmark, successfully completing 33.8% of tasks, compared to 27.8% for Haiku and only 17% for Luna.
Multimodal integration with Muse Wonder and expansion of the Muse model family
Ember is designed to work inside agent systems alongside Meta’s larger Muse Wonder model, with the company summarizing their division of labor: “Muse Wonder plans, Muse Ember handles the heavy lifting.” Ember accepts text, images, audio, and video, allowing agents to read screenshots, charts, and scanned documents, or work from voice memos and meeting and video recordings without a separate model. It is also the first model trained by Meta using its larger Muse Wonderful model.
Developers can use it within the Muse Code environment and via the Meta Models Developer API. Unlike earlier Muse releases, Ember’s model weights are not publicly available. Ember is the fourth model in the Muse family following the mid-tier Spark in April, its open-source counterpart Glimmer in August, and the flagship Wonder in October, running parallel to the launch of the Muse consumer agent in September and the addition of spaces and video chat features in early November.
The launch of Muse Ember reinforces Meta’s position in the fierce competition for the operational execution layer within enterprise agent ecosystems. As companies increasingly rely on workflows that invoke models thousands of times daily to process documents and multimedia, cost efficiency per completed task and context window capacity are becoming critical factors, no less important than the higher-level planning capabilities handled by larger flagship models.
Frequently asked questions
Question: What are the pricing and context window capacities for Muse Ember?
Answer: It costs $0.10 per million input tokens, $0.004 for cached input, and $0.50 per million output tokens, featuring a one-million-token context window and up to 512,000 output tokens.
Question: How does Muse Ember economically outperform competitors in reasoning tasks?
Answer: It costs about $0.14 per task on the Artificial Analysis Intelligence Index with a score of 36 points, compared to $0.41 for Claude Haiku 5.5 at the same score and $0.15 for Luna.
Question: What input types does Muse Ember support and how does it work with Muse Wonder?
Answer: It accepts text, images, audio, and video to analyze screenshots and recordings directly, serving as a high-density execution layer for tasks planned by the larger Muse Wonder model.
Question: Did Meta release Muse Ember with open-source weights for the public?
Answer: No; Muse Ember’s weights are closed and unavailable to the general public, though it is accessible to developers via the Muse Code tool and Meta’s cloud model API.