Meta introduced two powerful artificial intelligence models, Muse Wonder and Muse Wonder Full, on Monday, October 12, 2026, marking a major expansion of its efforts to compete with leading AI developers. The new systems are designed to handle complex assignments, remember information across conversations and complete lengthy tasks with less human supervision.
Developed by Meta Superintelligence Labs, the models represent the company's most advanced AI technology to date. They build on Muse Spark, an earlier model designed for speed and efficiency, but offer greater computing capabilities for demanding professional and personal tasks.
Muse Wonder is now available through the Meta Model API and Muse Code, the company's AI-powered development platform. The more powerful Muse Wonder Full is restricted to Muse Code subscribers and is not being offered through the API.
The launch follows more than a year of restructuring and investment under CEO Mark Zuckerberg, who has made advanced artificial intelligence a central priority for Meta's future.
Two Models Designed for Different Tasks
Meta developed the Muse Wonder family to handle assignments requiring extended reasoning, planning and the ability to work with different kinds of information.
Muse Wonder is intended for developers and businesses that need advanced capabilities through an API. It can assist with programming, analyze complicated documents and support automated software agents that perform multiple steps to accomplish a goal.
Muse Wonder Full is designed for more demanding workloads and provides additional capabilities for sophisticated coding and professional assignments. Meta describes it as its most capable model.
Both systems are aimed at what the industry calls agentic AI. Unlike a conventional chatbot that primarily responds to individual questions, an AI agent can plan a sequence of actions, use software tools and continue working toward a larger objective.
Meta provided examples involving insurance claims, financial documents, legal agreements and application development. A user could ask the system to examine a complicated lease, identify important provisions or help build a functioning software application.
These capabilities are intended to reduce the amount of manual work required to complete complicated projects. However, performance can vary depending on the assignment, available tools and the quality of information supplied.
The company is also preparing to integrate Muse Wonder into its Muse assistant platform. Meta said the expanded capabilities will initially become available to subscribers before reaching a wider audience.
Active Memory Helps AI Remember Important Details
One of the most important features introduced with the new models is Active Memory.
The technology is designed to preserve useful information from earlier interactions and retrieve it when relevant to a later assignment. This could allow an assistant to continue working on a long-running project without requiring users to repeatedly explain the same background information.
For example, someone managing a business project might share objectives, deadlines and previous decisions during several conversations. An assistant with effective memory could use those details when preparing later documents or reviewing progress.
Meta also introduced Active Focus, a feature intended to help the models continue pursuing difficult goals.
When an approach fails, the system can attempt another method, recover from errors and continue working. Meta says the models are designed to request clarification when necessary and seek confirmation before taking important actions.
Together, these features are intended to make AI assistants more useful for work that extends beyond a single conversation.
TechCrunch reported that Meta placed particular emphasis on a benchmark called MemoryBench, which evaluates how effectively AI systems retain and use information over time.
According to Meta's internal testing, Muse Wonder achieved stronger results on that benchmark than several competing systems, including models from OpenAI and Anthropic.
Research scientist Kavya Kopparapu explained during the product announcement that the company had invested considerable effort in developing systems capable of remembering important information across extended periods.
Those results support Meta's claims about improved memory performance, although company-controlled benchmarks do not establish that the models will outperform competitors in every real-world situation.
Independent Tests Show Significant Improvements
Independent testing provides additional evidence that Meta has made progress with the Wonder models.
Artificial Analysis, a company that evaluates AI systems, reported that Muse Wonder achieved a score of 52 on its Intelligence Index, while Muse Wonder Full scored 56. Both were tested at their highest reasoning settings.
The earlier Muse Spark 1.3 model scored 48 under comparable conditions.
That represents a four-point improvement for Muse Wonder and an eight-point improvement for Muse Wonder Full.
The independent evaluation also identified gains in software engineering, professional assignments and tasks requiring automated agents.
Muse Wonder Full improved across all ten evaluations included in the Artificial Analysis Intelligence Index compared with Muse Spark 1.3.
Its Terminal-Bench 4.0 result increased from 33 percent to 57 percent, showing stronger performance on technical tasks involving command-line tools and software environments.
The findings suggest that Meta has made measurable progress rather than relying entirely on its own performance claims.
However, the results do not establish that Muse Wonder Full is the strongest AI system overall. Competing models remain ahead on some benchmarks, and different testing methods can produce different rankings.
Meta has also published evaluations covering legal reasoning, financial analysis, health-related questions and extended software development.
Those measurements help demonstrate the systems' capabilities, but benchmark results should not be interpreted as guarantees of accuracy. Professional applications involving health, law or financial decisions still require appropriate human review.
Muse Wonder Pricing and Availability
Meta is introducing two different access models for the new systems.
Muse Wonder is available through the Meta Model API, allowing developers to incorporate its capabilities into software applications and business systems.
The company lists API pricing at $4 per million input tokens and $20 per million output tokens. Cached input costs $0.20 per million tokens.
Tokens are the units AI systems use to process text and other information. Developers generally pay according to how many tokens their applications send to and receive from a model.
Both Wonder models support a context window of approximately one million tokens, allowing them to work with substantial amounts of information during a single session.
Muse Wonder Full follows a different commercial approach. It is available through Muse Code subscriptions rather than direct API access.
Meta offers access through its Power Usage plan at $50 per month and Maximum Usage plan at $200 per month.
Restricting the more capable model to subscribers gives Meta greater control over how its most computationally demanding technology is used.
It also creates a distinction between developers who want API access and customers who prefer working directly within Meta's programming tools.
The company has not announced a date for making Muse Wonder Full available through a public API.
Meta's AI Restructuring Begins to Produce Results
The Wonder release follows a substantial reorganization of Meta's artificial intelligence operations.
Zuckerberg began reshaping the company's AI strategy in 2025, recruiting researchers and reorganizing development teams to accelerate progress toward more capable systems.
Alexandr Wang, founder of Scale AI, became a leading figure in Meta's new AI organization as the company increased investment in research talent and computing infrastructure.
Muse Spark emerged as an earlier product of that effort. The Wonder family represents a move toward larger models intended to compete with the most capable systems developed by OpenAI, Anthropic and Google.
The models were trained using infrastructure at Meta's Prometheus data center in central Ohio, which became operational earlier in 2026.
Advanced AI development requires enormous computing capacity, including specialized processors, power infrastructure and systems capable of handling extensive training workloads.
Meta's investment in these facilities reflects its belief that stronger AI models will become increasingly important to its consumer products and business services.
The company is also seeking opportunities beyond its established advertising operations, including subscriptions and developer services.
Reuters reported that Meta shares rose approximately 4 percent during afternoon trading following the announcement, reflecting investor interest in the company's latest technological development.
Privacy and Safety Remain Important Questions
The introduction of stronger memory and agent capabilities creates additional responsibilities for Meta.
An assistant that can access messages, calendars, financial records or business documents may become more useful, but it also handles information that requires careful protection.
Meta says it has trained the Wonder family to avoid revealing private information without permission and to decline requests involving harmful activities.
The company also says its systems are designed to help developers identify security weaknesses while maintaining protections against misuse.
For tasks involving potentially significant actions, the models are intended to seek user confirmation before proceeding.
These safeguards are important because an agent that can interact with external software may cause problems if it misunderstands instructions or performs an unauthorized action.
Meta says it evaluates its most capable models under the Meta Superintelligence Scaling Framework and plans to publish additional safety assessments.
The effectiveness of those protections will become clearer as developers and subscribers use the models in more varied situations.
A New Competitor in Advanced AI
The release places Meta in a stronger position within the market for advanced AI systems.
The company now offers models designed for coding, complex reasoning, long-running assignments and personal assistance, supported by a growing ecosystem of development tools.
Independent evaluations indicate meaningful improvements over Meta's earlier technology, although competitors continue to lead in some areas.
Meta's longer-term strategy focuses on what Zuckerberg calls personal superintelligence, a vision of AI assistants capable of helping individuals manage increasingly complicated activities.
For now, Muse Wonder is available to developers through the Meta Model API and Muse Code. Muse Wonder Full is available to eligible Muse Code subscribers.
Meta's next announced step is to bring Wonder-powered capabilities to the Muse assistant, beginning with subscribers. The company has not provided a specific date for that wider rollout.