Meta launched the Muse Ember AI model on Nov. 9 as the company’s smallest and most cost-efficient Muse system. Developed by Meta Superintelligence Labs, it targets repetitive tasks AI agents perform, including coding, tool calls, and information extraction.
Ember costs $0.10 per million input tokens and $0.50 per million output tokens. Meta says its performance compares favorably with Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna at lower operating costs.
Meta Muse Ember AI Model Focuses on Lower Costs
Meta designed Ember to work alongside larger models, not replace them. Muse Wonder can handle complex planning, while Ember executes routine tasks at scale.
For example, a customer support system could use a larger model to establish response rules. Ember could then apply those rules across thousands of customer requests.
Meta’s comparisons use the Artificial Analysis Intelligence Index, which evaluates reasoning, coding and other AI capabilities.
The company reports an Intelligence Index score of 36 for Ember, matching Claude Haiku 5.5’s highest score. Meta estimates that Ember achieves comparable performance at roughly one-third of Haiku’s cost.
These comparisons depend on benchmark settings and pricing assumptions. They do not guarantee identical results in production applications.
Multimodal Support and Agent Capabilities
Ember supports text, images, audio and video inputs. Agents can analyze screenshots, scanned documents, charts and recordings without switching to another model.
The model also handles coding subtasks such as reading files, running tests and making targeted changes. Its multimodal capabilities extend to computer interaction and document processing.
Meta reports that Ember scored 61.6% on Agents’ Last Exam, compared with 57% for Claude Haiku 5.5. GPT-6 Luna scored 57.1% in the same comparison.
However, Ember did not lead every evaluation. Claude Haiku 5.5 performed better on FrontierCode 2.0, which tests complex engineering tasks.
Muse Ember Pricing and Availability
Developers can access Ember through the Meta Model API and Muse Code using the identifier muse-ember-1.0.
The model offers a 1 million-token context window and supports up to 512,000 output tokens. Cached input tokens cost $0.004 per million.
Ember joins Muse Spark, Glimmer and Muse Wonder in Meta’s growing AI model portfolio. It is also the first Muse model trained using Muse Wonder Full.
Before release, Meta evaluated Ember for cybersecurity, biological risks and potentially unsafe agent behavior. The company said the model remained below its high-risk thresholds.
The release gives developers another option for managing frequent AI agent tasks. Its main selling point is lower execution costs without relying on a large model for every operation.
