Meta Unveils Muse Wonder AI Models to Challenge OpenAI and Anthropic

Meta has announced two new artificial intelligence models, Muse Wonder and Muse Wonder Full, as part of its effort to become a stronger competitor in the rapidly evolving AI industry. Announced on October 12, 2026, the models are designed to handle complex personal and professional tasks, including software development, financial analysis, legal document interpretation and healthcare-related information processing. Meta described the release as a major milestone following approximately 14 months of restructuring and rebuilding its AI operations. The new models build on the foundations of Muse Spark, the company’s earlier AI model, and are intended to move Meta closer to its ambition of developing highly capable personal AI assistants. The announcement also signals Meta’s determination to compete directly with leading AI developers such as OpenAI and Anthropic.

One of the most important features of the Muse Wonder models is their ability to work on tasks that require several steps rather than simply answering individual questions. Meta introduced two capabilities called Active Memory and Active Focus. Active Memory allows the models to retain relevant information across conversations and bring it back when needed, while Active Focus helps them continue working towards a goal when they encounter difficulties. According to Meta, the models can try alternative approaches, recover from errors and request clarification when instructions are unclear. They are designed to support activities such as understanding an insurance claim, reviewing a rental agreement, analysing financial information and building software applications. These capabilities could make AI assistants more useful for people who need help managing complex projects over extended periods.

The company is also positioning Muse Wonder as a major step forward in AI-powered software development. The models are designed to generate code, work with development tools and help users build applications ranging from simple games to more sophisticated digital products. Meta says the new systems achieve strong results across several professional and personal agent benchmarks, including evaluations involving coding, legal work and financial tasks. The more powerful Muse Wonder Full is intended for demanding workloads, while the standard Muse Wonder provides developers with access to the new model through Meta’s API and Muse Code. Although Meta has presented encouraging benchmark results, independent testing will be important for determining how consistently the models perform against competing systems in real-world conditions.

Safety and privacy are also central to the launch because the models are designed to operate as AI agents capable of handling sensitive information and carrying out actions on users’ behalf. Meta says it has trained Muse Wonder to avoid sharing private information without permission, reject requests to facilitate harmful activities and strengthen cybersecurity defences. The company is also attempting to preserve useful defensive cybersecurity capabilities, allowing developers and security professionals to identify vulnerabilities and improve their software. These safeguards are particularly important as AI agents gain access to documents, messages, calendars and other connected services. However, Meta has said that a more detailed safety evaluation report will be published separately, meaning the public will need further evidence to assess the effectiveness of the new models’ protections.

Meta has begun making Muse Wonder available to developers through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. The company plans to integrate Muse Wonder into its consumer-facing Muse assistant, initially rolling it out to subscribers before expanding availability. The models are expected to support more advanced experiences, including generating digital content and completing tasks that require greater reasoning ability. The launch comes amid intense competition between technology companies investing billions of dollars in AI infrastructure, model training and autonomous agents. For Meta, the success of Muse Wonder will depend not only on benchmark performance but also on reliability, user trust, developer adoption and the ability to deliver useful AI features across its wider ecosystem. If the models perform as intended, they could strengthen Meta’s position in the race to build AI assistants that do more than generate text and can instead complete meaningful work.