Meta has officially launched its Muse Wonder and Muse Wonder Full AI models, marking a definitive shift in the company’s artificial intelligence strategy. Released on October 12, 2026, these frontier-class models represent a transition away from the open-weights philosophy of the Llama series toward a proprietary ecosystem managed by Meta Superintelligence Labs.
The new models are the first major release under the leadership of Chief AI Officer Alexandr Wang, following Meta’s $14.3 billion acquisition of Scale AI. Unlike previous releases, Meta is positioning the Muse family as a tiered, revenue-generating product line designed to compete directly with high-end offerings from Anthropic and OpenAI.
Tiered Access and Pricing Structure
The rollout distinguishes between the two models based on performance and accessibility. The standard Muse Wonder model is available through the Meta Model API for general development and is also integrated into the Muse Code terminal agent. For API users, Meta has set pricing at $4 per million input tokens and $20 per million output tokens.
The more powerful Muse Wonder Full is currently restricted to paid Muse Code subscribers on the “Power” and “Maximum” plans. This exclusivity suggests a strategic focus on capturing enterprise and high-end developer markets rather than the broad-access model Meta previously championed with its open-source initiatives.
Technical Capabilities and “Active” Features
Both models in the Muse family feature a context window of approximately 1.05 million tokens, allowing for the processing of vast datasets and long-form codebases. Meta has introduced two specific architectural features aimed at improving agentic workflows:
Active Memory: This capability allows the model to maintain context across multiple, separate conversations, reducing the need for constant re-prompting or redundant data input.
Active Focus: This feature is designed for sustained agentic goal tracking, allowing the AI to maintain a specific objective over long execution cycles without drifting from the original task requirements.
The training of these models was conducted at Meta’s “Prometheus” data center, a gigawatt-scale facility built specifically to support frontier-class intelligence.
Benchmark Performance
Independent evaluations indicate that Meta has successfully closed the gap with established frontier models. According to the Artificial Analysis Intelligence Index, Muse Wonder Full achieved a score of 56. This performance places it ahead of GPT-6 Astra, which currently holds a score of 53. The standard Muse Wonder model remains competitive with a score of 52.
These benchmarks suggest that Meta’s proprietary approach, fueled by the data labeling expertise brought in by the Scale AI acquisition, has yielded a model family capable of handling complex reasoning and coding tasks at the same level as the industry’s top-tier proprietary systems. By moving Muse Wonder Full behind a subscription wall, Meta is signaling a new phase where its most advanced research is reserved for its highest-paying users.
