Meta is emphasizing the relationship between performance and operating costs. According to results presented by the company using Artificial Analysis benchmarks, Muse Ember offers stronger performance per dollar than Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna across the reasoning configurations compared.
At its highest reasoning setting, Ember achieved an Artificial Analysis Intelligence Index score of 36, matching Haiku's top score while costing substantially less per task. Meta estimates Ember's cost at approximately $0.14 per benchmark task, compared with $2.39 for Muse Spark 1.3.
Those comparisons use Meta's published pricing and benchmark results, rather than measuring every possible real-world workload. Performance also varies by evaluation, with Claude Haiku outperforming Ember on some engineering tasks.
The pricing structure is intended to make Ember practical for applications that repeatedly process large amounts of information. In addition to its $0.10 input-token rate, Meta charges $0.50 per million output tokens and $0.004 per million cached input tokens.
Ember supports a context window of one million tokens and can generate up to 512,000 output tokens. Those limits allow agents to work with extensive documents or code repositories without splitting the material across numerous requests.
One of Meta's primary use cases involves combining Ember with its more powerful Muse Wonder model. In this arrangement, Wonder can handle difficult planning and decision-making while Ember executes the numerous smaller operations required to complete a workflow.
For example, an enterprise AI agent could use a larger model to establish how a customer-service issue should be handled, then delegate repeated processing tasks to Ember across thousands of cases.
Meta summarizes the approach with the phrase, “Muse Wonder plans, Muse Ember handles the volume.”
The company also designed Ember to operate across multiple types of information. The model can process text, images, audio, and video natively, enabling agents to analyze screenshots, interpret charts, read scanned documents, and work with recordings without switching to a separate model.
These capabilities extend to computer-based tasks in which an agent must interpret visual information and interact with applications. Meta reports that Ember scored 69.5% on OSWorld 2.1, an evaluation of computer-use capabilities, compared with 68.9% for Haiku 5.5 and 50.5% for Luna at their highest reasoning settings.
On AutomationBench-AA, which evaluates end-to-end business processes, Ember achieved 37.8%, ahead of Luna's 20.2% and Haiku's 13.9%. It also scored 61.6% on Agents' Last Exam, compared with 57.1% for Luna and 57.0% for Haiku.
Coding performance showed a more mixed picture. Ember reached 40.9% on Terminal-Bench 4.0, slightly ahead of Haiku's 39.2%, but Haiku performed better on FrontierCode 2.0, scoring 34.9% against Ember's 27.2%.
Meta says early enterprise testing has focused on tasks involving document analysis, coding, and structured business workflows. Companies involved in testing include Box, Cline, Oracle, Cloudflare, and You.com.
Box reported promising results when using Ember to evaluate contracts against negotiation guidelines and assess job candidates against detailed requirements. The company highlighted the model's ability to combine accuracy with efficient token usage for enterprise content workflows.
The release also includes safety evaluations under Meta's Superintelligence Scaling Framework. According to the company, Ember remained below its designated high-risk thresholds in assessments covering chemical and biological threats, cybersecurity, and potential loss of control.
Meta additionally tested the model's behavior when operating on behalf of users, including its handling of sensitive information, resistance to malicious instructions embedded in external content, and ability to seek human input before consequential actions.
Muse Ember is available now through Muse Code and the Meta Model API under the identifier muse-ember-1.0, with support for the Responses API.
The launch follows several additions to Meta's Muse lineup this year, including Muse Spark in April, the open-source Glimmer model in August, and Muse Wonder in October. Meta also introduced its consumer-facing Muse Agent in September and recently added features including Spaces and video chat.
With Ember, Meta is extending that model portfolio toward applications where the cost of repeated AI operations is a central consideration. Rather than competing exclusively on the capabilities of its largest models, the company is offering a smaller system designed to execute everyday agent tasks at scale while reserving more powerful models for work requiring deeper reasoning.
This analysis is based on reporting from Meta & TechCrunch.
Image courtesy of Meta.
This article was generated with AI assistance and reviewed for accuracy and quality.
