Meta Releases Muse Ember Lightweight Reasoning Model

To put this into perspective, Claude Haiku 5.5 requires approximately $0.248 per task to achieve a comparable score of 34.3. Effectively, Muse Ember matches Haiku's best performance at roughly one-third of the cost. Additionally, Ember outperforms every reasoning setting of GPT-6 Luna in intelligence-per-dollar comparisons, marking a significant shift in the efficiency frontier.

These benchmarks were evaluated using a specialized methodology focused on 'reasoning effort settings,' measuring how much intelligence a model provides relative to the computational cost required to complete a specific, repeatable task. This approach is more relevant to agent developers than traditional MMLU scores, as it reflects real-world operational economics.

• Scores 35.8 on intelligence-vs-cost-per-task at $0.136
• Matches Claude Haiku 5.5 performance at ~33% of the cost
• Outperforms GPT-6 Luna in intelligence-per-dollar efficiency
• Evaluated on task-specific reasoning cost rather than just raw accuracy

API Pricing

Meta has introduced a highly aggressive pricing structure for Muse Ember, aimed at making large-scale agent deployment economically viable. The proposed pricing is designed to undercut the 'small model' tier of major competitors, providing a massive runway for startups building agent-heavy applications.

The cost structure is straightforward, focusing on high-volume input and output efficiency. This makes it particularly attractive for tasks like long-document parsing where input tokens are abundant, or for iterative coding subtasks where output generation is frequent.

• Input Price: $0.10 per million tokens
• Output Price: $0.50 per million tokens
• Targeting high-volume, low-margin agentic workflows

Use Cases

Given its architecture and pricing, Muse Ember is best suited for specific, high-frequency roles within an AI ecosystem. It is not a replacement for a 'reasoning giant' like GPT-6, but rather the workhorse that executes the granular steps those giants delegate.

Primary use cases include autonomous coding agents (handling subtasks like unit test generation or refactoring), complex routing (deciding which model to call next), and massive-scale data extraction from unstructured documents. Its 1M context window also makes it a premier choice for advanced RAG implementations where entire datasets must be analyzed in a single pass.

• Agentic Tool Calling & Function Execution
• Coding Subtasks & Automated Refactoring
• Long-form Document Parsing & Extraction
• High-throughput Routing & Classification
• Large-scale RAG (Retrieval-Augmented Generation)

Getting Started

Developers can begin integrating Muse Ember into their workflows immediately. The model is available through the Meta Model API, providing a standard interface for seamless integration into existing Python or JavaScript-based agent frameworks.

For developers working within specialized environments, Ember is also available via Muse Code, Meta's dedicated suite for AI-assisted development. This provides a direct path for engineers looking to leverage the model's efficiency for automated software engineering tasks.

• Access via Meta Model API
• Integrated availability in Muse Code
• Compatible with standard agentic frameworks

API Pricing — Input: $0.10 per million tokens (proposed) / Output: $0.50 per million tokens (proposed) / Context: 1M tokens

Sources