Meta has unveiled two new artificial intelligence models — Muse Wonder and Muse Wonder Full. The company says they are capable of solving complex tasks at the level of leading competing systems, including programming. Among the main improvements is the ability to retain and use information about the user over long periods, which Meta calls active memory.

The new models were developed by Superintelligence Lab, Meta’s division focused on creating advanced AI systems. Unlike the more compact Muse Spark, which is geared toward fast responses, Muse Wonder is designed for more complex tasks and delivers stronger results in a number of tests.

On par with the market’s leading models

According to Meta, the benchmark results of Muse Wonder and Muse Wonder Full are comparable to those of Fable 5.1 and GPT-6 Astra. The company specifically highlights improvements in programming and handling complex prompts. These results are presented in the technical documentation published alongside the models, although for now they are based on the developer’s own tests.

Meta describes the new models as the next stage in scaling its AI systems and a step toward “personal superintelligence” — an assistant that does not simply answer individual questions, but takes into account a person’s goals, habits, and interaction history.

Access to Muse Wonder through the Meta Model API will cost $4 per million input tokens and $20 per million output tokens. In terms of price, the model is comparable to GPT-6 Sol and Opus 5.5. The more powerful Muse Wonder Full will not be available in the API: only Muse Code subscribers will be able to use it. Both models will also be used to power the Muse AI agent.

AI that remembers the user for months

One of the key features of Meta’s new models is what it calls active memory. It is meant to help the AI retain important details from previous conversations and combine them, even if days, weeks, or months pass between interactions.

For a personal assistant, this means the ability to take into account not only the content of the current request, but also the accumulated context. For example, information about the user’s preferences, plans, and current tasks may prove useful in later conversations, sparing the person from having to explain the situation from scratch every time.

As evidence of these improvements, Meta cites results from MemoryBench, a test that evaluates approaches to continual learning and memory building. According to the company, Muse Wonder outperforms Fable, Opus, and Astra on it.

“We put tremendous effort into making sure the model can remember what matters to you across days, weeks, months, and beyond,” Meta researcher Kavya Kopparapu said in a video accompanying the launch.

At the same time, benchmark results alone do not guarantee that the model will always retrieve the right information correctly in real-world interactions. For users, another important question remains what data the assistant stores, how long it keeps it, and how easy it is to manage that information.

The first major step after the AI division’s restructuring

This is Meta’s first release of a flagship model since the major reorganization of its AI division led by Scale AI founder Alexandr Wang in August 2025. Since then, the company has introduced the more compact Muse Spark and the Muse AI agent, but had not released models it positioned at the level of competitors’ most advanced systems.

The new models were trained at the Prometheus data center in central Ohio, which began operations in 2026.

The launch of Muse Wonder shows that Meta aims to compete not only through affordable models and mass-market AI products, but also in the segment of the most powerful systems. At the same time, the company places special emphasis on personalization: its goal is to create an assistant that can do more than simply carry out a one-off request, instead taking into account the history of interactions with a person and helping with long-term tasks.