Meta Launches Muse Ember as Its Smallest and Most Efficient AI Model

Meta has introduced Muse Ember, its smallest and most efficient artificial intelligence model, claiming it delivers better performance per dollar than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 across several reasoning and agent-based tasks. The release marks a significant expansion of the company AI model lineup, joining existing systems like Muse Spark and Muse Wonder in enterprise environments. Industry observers note that the efficiency frontier of artificial intelligence continues to shift rapidly as developers demand higher performance at lower operating expenditures.

The new model focuses heavily on complex reasoning and repetitive artificial intelligence agent operations, including coding, tool calls, task routing, information extraction, and processing lengthy documents. According to Meta, Muse Ember delivers performance comparable to Claude Haiku 5.5 at roughly one-third the cost in certain configurations, while outperforming GPT-6 Luna across the tested reasoning settings. The system is designed to integrate smoothly into automated workflows where speed and affordability are paramount for business efficiency.

Performance and Benchmarks on Leading Industry Tests

Meta states that Muse Ember sits directly on the efficiency frontier of the Artificial Analysis leaderboard, meaning no tested model achieves a higher intelligence score at a lower cost per task. At maximum reasoning effort, Muse Ember scored 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 and exceeding GPT-6 Luna's score of 28. The model also performed well in agent-based tasks, scoring 37.8 percent on AutomationBench-AA, compared with 20.2 percent for GPT-6 Luna and 13.9 percent for Claude Haiku 5.5.

Further evaluation reveals strong capabilities across multiple technical domains, though competitors maintain advantages in specific niches. Muse Ember scored 40.9 percent on Terminal-Bench 4.0, finishing ahead of Claude Haiku 5.5 at 39.2 percent and GPT-6 Luna at 16.4 percent. However, Claude Haiku 5.5 performed better on FrontierCode 2.0-AA, scoring 34.9 percent compared with Muse Ember's 27.2 percent. Additional testing figures demonstrate competitive results in the Personal Superintelligence Index, Coding Agent Index-AA, and OSWorld 2.1 benchmarks.

In the Personal Superintelligence Index, Muse Ember reached 46.8 percent, outperforming GPT-6 Luna at 37.0 percent and Claude Haiku 5.5 at 36.1 percent. On the Coding Agent Index-AA, Muse Ember achieved 46 points, surpassing Claude Haiku at 37 and GPT-6 Luna at 31. For computer-based operational tasks in OSWorld 2.1, the model scored 69.5 percent, closely contested by Claude Haiku at 68.9 percent and well above GPT-6 Luna at 50.5 percent. PerceptionBench testing placed Muse Ember at 80.6 percent, compared to 78.6 percent for Claude Haiku and 62.8 percent for GPT-6 Luna.

Core Features and Capabilities of the New Architecture

Muse Ember supports text, images, audio, and video, allowing it to analyze screenshots, charts, scanned documents, recordings, and other multimodal content. The model features a 1-million-token context window and supports outputs of up to 512,000 tokens, allowing it to process large codebases and document collections without losing contextual continuity. It can perform computer-based tasks, review contracts, analyze documents, run coding tests, and complete workflows involving multiple connected applications.

Meta suggests developers can pair Muse Ember with Muse Wonder, using the larger model for complex planning while Ember handles repetitive tasks efficiently. The model is accessible through the Responses API and integrates directly into development environments like Muse Code. Developers can deploy the system via the Meta Model API, with enterprise cloud infrastructure support provided by Google Cloud and Oracle Cloud platforms.

Pricing, Token Costs, and Comparison with Muse Spark 1.3

Muse Ember is available through Muse Code and the Meta Model API under the designation muse-ember-1.0. Meta has established competitive pricing tiers for input and output token usage, aiming to drive wider adoption among cost-conscious enterprises and independent software developers. Input tokens cost $0.10 per 1 million tokens, while cached input tokens are priced at $0.004 per 1 million tokens. Output tokens are billed at $0.50 per 1 million tokens, resulting in an estimated cost of approximately $0.14 per Artificial Analysis task.

According to Meta, Muse Ember costs approximately $0.14 per Artificial Analysis task, compared with $2.39 for Muse Spark 1.3. Despite the massive difference in operating expenditure, the two models scored 36 and 37, respectively, on the Artificial Analysis Intelligence Index. Muse Ember is designed specifically for repetitive coding and agent-based tasks, while Muse Spark 1.3 targets more demanding coding and agentic workflows that require deeper computational resources.

Safety Evaluations and Future Development Plans

Meta states it evaluated Muse Ember under its Superintelligence Scaling Framework, covering cybersecurity, chemical and biological risks, and potential loss-of-control scenarios. The company reported that the model remained safely below its high-risk thresholds across all categories tested during the evaluation phase. Meta also assessed whether Muse Ember follows user instructions, protects sensitive information, resists malicious instructions embedded in external content, and requests human approval before executing high-impact actions.

Muse Ember stands as the first model Meta trained using Muse Wonder Full, establishing a new baseline for future efficiency-focused iterations. It joins Muse Spark and Muse Wonder in the company growing commercial artificial intelligence model lineup. Meta confirms that all three models are now available on Google Cloud and Oracle Cloud, with additional efficiency-focused models currently in active development for future deployment.