Meta Releases Muse Ember, a Lightweight Model Built for Agent Execution

Meta Superintelligence Lab introduced Muse Ember on November 9, 2026. The system is the lab's smallest and most budget-friendly offering to date, designed explicitly for multi-step agent routines such as tool execution, code generation, routing, and data extraction. Available immediately via the Meta Model API, the model costs $0.10 per million input tokens and $0.50 per million output tokens.

Meta emphasized Ember's cost efficiency relative to other small models on the market. According to Artificial Analysis benchmarks shared by the company, Ember matched the peak performance of Anthropic's Claude Haiku 5.5 at roughly one-third the cost per task, while outperforming OpenAI's GPT-6 Luna across all reasoning settings. On the AutomationBench-AA benchmark, Ember scored 37.8 percent, compared to 20.2 percent for GPT-6 Luna and 13.9 percent for Haiku 5.5.

The system supports a 1-million-token context window alongside an output capacity of up to 512,000 tokens. Ember is trained to process text, image, and audio inputs natively, allowing autonomous systems to review complex forms, code repositories, screenshots, and audio recordings without separate processors. The release follows earlier Muse models including Spark, Glimmer, and the flagship Wonder system.

As software agents take on larger corporate workloads, the cost of running thousands of background steps becomes a major bottleneck. Lightweight models like Ember allow companies to deploy autonomous workflows at scale without incurring the massive token bills of flagship models.

The release also reinforces an emerging division of labor in software engineering: larger models like Muse Wonder handle high-level reasoning and planning, while smaller, specialized systems handle repetitive execution. This setup lets enterprises tackle complex business processes reliably while keeping operating expenses sustainable.