Meta's Muse Ember Targets Budget-Conscious AI Agent Deployments

Meta Platforms Inc. has launched Muse Ember, a budget-friendly artificial intelligence model designed to undercut rivals on price. The model represents the first system Meta trained using Muse Wonder, the company's leading large language model, though Meta has not disclosed which specific training phase benefited from this approach.

Post-training—the refinement of model capabilities through additional data—represents one likely application of Muse Wonder in Muse Ember's development. Recent work by Thinking Machines Labs Inc. demonstrated that smaller language models can execute full post-training workflows independently. Anthropic PBC separately announced last week that its models may soon gain capabilities to substantially speed up machine learning research efforts.

In typical deployment scenarios, frontier models like Muse Wonder function as coordinators within AI agent networks. When users submit requests, these flagship models decompose them into discrete subtasks and route them to specialized agents built on lower-cost alternatives. Meta engineered Muse Ember specifically for this subordinate agent role.

Meta benchmarked Muse Ember using the Artificial Analysis Index, which combines nine performance metrics across domains including software development and financial analysis. According to Meta's testing, Muse Ember achieved parity with Anthropic's Claude Haiku 5.5 entry-level model while costing roughly one-third as much. The model also surpassed OpenAI Group PBC's GPT-6 Luna offering.

In a separate evaluation across nine additional benchmarks, Meta ran Muse Ember, Claude Haiku 5.5 and GPT-6 Luna in "max" mode, allowing each to utilize full hardware resources rather than optimizing for cost. Under these conditions, Muse Ember outperformed both competitors on seven of the nine tests.

Business process automation emerged as Muse Ember's strongest area of differentiation. While most benchmark advantages measured only single-digit percentage gains over Haiku and Luna, Muse Ember achieved a 17% lead on AutomationBench-AA, a suite of 657 tasks spanning widely-used business platforms including Gmail and Salesforce.

Muse Ember supports a 1 million token context window, permitting ingestion of extensive business documents. The model accepts images, audio and video within user prompts and can generate responses extending to 512,000 tokens.

Through Meta's Model API, Muse Ember carries pricing of 10 cents per million input tokens and 50 cents per million output tokens. An integrated caching mechanism can recycle frequently-used prompts to eliminate redundant computation. When caching is engaged, input token pricing decreases to 0.4 cents per million tokens.