Meta has introduced two new artificial intelligence models, Muse Wonder and Muse Wonder Full, marking a major step in the company’s efforts to compete in the advanced AI market. Developed by Meta’s Superintelligence Lab, the models are designed to handle more demanding tasks, including complex coding and information-intensive work.
The launch represents a shift from the company’s earlier focus on lightweight AI models. While Muse Spark was designed to deliver responses with low latency, the new Wonder models are larger and aim to provide stronger performance on challenging tasks.
Meta says the models achieve benchmark results comparable to leading AI systems, including Fable 5.1 and GPT-6 Astra. The company describes the release as another step toward its broader goal of developing personal AI assistants that can understand users and help them over extended periods.
Muse Wonder Models Aim to Deliver More Advanced AI Performance
Meta has positioned Muse Wonder and Muse Wonder Full as more capable models than its earlier lightweight offering.
According to the company’s model documentation, the new models perform well on several benchmarks and are designed to handle tasks that require more advanced reasoning and coding capabilities.
The two models also have different access arrangements. Muse Wonder will be available through the Meta Model API, allowing developers to integrate it into their applications. Muse Wonder Full, meanwhile, will be restricted to Muse Code subscribers and will not be available through the API.
Both models are also expected to power Meta’s Muse agent, connecting the new AI capabilities with the company’s broader efforts to develop AI systems that can assist users with tasks.
Meta Introduces New Pricing for Muse Wonder
The launch includes a pricing structure for developers who want to access Muse Wonder through the Meta Model API.
According to the announced pricing, Muse Wonder will cost:
- Input tokens: $4 per million tokens
- Output tokens: $20 per million tokens
Input tokens generally represent the text or other supported information sent to a model, while output tokens represent the content generated in response.
The pricing makes the cost of using the model an important consideration for developers building AI-powered applications. The final expense will depend on how much information an application sends to the model and how many tokens it generates.
Muse Wonder Full will follow a different distribution model, with access limited to Muse Code subscribers rather than API customers.
Active Memory Is a Major Focus of Meta’s New AI Models
One of the most notable features highlighted by Meta is what the company calls active memory.
The idea is to help an AI model retain and connect relevant information shared by a user over time. Instead of treating every interaction as an isolated conversation, an assistant with effective memory could use information from earlier discussions to provide more relevant responses later.
For example, a personal AI assistant might remember a user’s preferred writing style, ongoing project requirements or recurring tasks. That information could help the assistant provide more consistent assistance across multiple sessions.
Meta has highlighted memory performance as an important area of development for its new models. According to the company, Muse Wonder achieved stronger results on MemoryBench than several competing models, including Fable, Opus and Astra.
MemoryBench evaluates approaches to memory construction and continual learning. Such benchmarks can help researchers compare how AI systems retain and use information, although performance on a benchmark does not automatically guarantee the same results in every real-world situation.
Research scientist Kavya Kopparapu also highlighted the work involved in helping models retain important information over days, weeks and months.
Why AI Memory Matters for Personal Assistants
AI memory could become increasingly important as companies move beyond chatbots toward assistants that help users manage ongoing activities.
A conventional chatbot is often used to answer individual questions, explain concepts or generate content. An assistant with longer-term memory could potentially build on previous interactions, track project context and reduce the need for users to repeat the same information.
This capability could be useful in software development, research, planning and workplace productivity. Developers could also explore applications that rely on a more consistent understanding of user preferences and previous tasks.
However, long-term memory introduces additional challenges. AI companies must consider how personal information is stored, how users can manage or delete it, and how systems avoid relying on outdated or incorrect information.
As AI assistants become more personalised, the balance between convenience, reliability and privacy will remain an important consideration.
Meta Steps Up Its Competition in the Frontier AI Market
The release of Muse Wonder and Wonder Full marks a new phase in Meta’s AI development strategy.
Following changes to its AI organisation under Alexandr Wang, Meta had released its lightweight Muse Spark model and introduced the Muse agent. However, the company had not made a comparable frontier-model release of the kind associated with leading AI developers.
The new models indicate that Meta is placing greater emphasis on performance, advanced capabilities and more sophisticated AI experiences.
The frontier AI market is becoming increasingly competitive, with technology companies investing in models that can handle complex reasoning, coding, research and multi-step tasks. These capabilities are also important for building AI agents that can do more than respond to individual prompts.
For Meta, stronger models could support its wider AI ecosystem and help the company compete for developers and users seeking more capable assistants.
What the Launch Means for Developers and Users
The release creates different opportunities depending on how people use AI.
Developers will be able to explore Muse Wonder through the Meta Model API, subject to the platform’s access requirements and pricing. This could allow them to test the model in coding tools, productivity applications and other AI-powered services.
Muse Code subscribers will have access to Muse Wonder Full under the announced subscription arrangement. Meanwhile, users of Meta’s Muse agent could benefit from the capabilities of both models as they are integrated into the assistant.
For everyday users, the biggest development may be the focus on memory and continuity. If these capabilities work reliably in practical use, AI assistants could become more useful for long-running projects and tasks that require context from previous conversations.
However, actual performance will depend on how the models behave outside benchmark tests, how well their memory features work in practice and how Meta implements privacy and user controls.
What Comes Next?
Meta’s latest models suggest that the company wants to compete more directly in the advanced AI market rather than focusing only on smaller, faster models.
The combination of stronger coding capabilities, improved benchmark performance and active memory could help Meta develop more capable personal assistants and AI agents.
The next important test will be how these models perform in real-world applications and whether developers and users find enough value in their capabilities to adopt them.
Conclusion
Muse Wonder and Muse Wonder Full represent a significant step in Meta’s AI strategy. By combining advanced model capabilities with a focus on long-term memory, Meta is working toward AI assistants that can do more than answer questions. Their practical success will depend on reliability, accessibility, privacy and the quality of the experience they deliver to users.