Meta Releases Muse Wonder to Reclaim AI Frontier

Meta has launched its latest artificial intelligence offerings, Muse Wonder and Muse Wonder Full, making them available via an API and through its software development tool, Muse Code. The top-tier Muse Wonder Full ranks directly below Anthropic's Claude Opus 5.5 and ahead of Claude Fable 5.1 on the Artificial Analysis Intelligence Index. Meta states that its new model also surpasses OpenAI's GPT-6 Astra and both Anthropic models on specialized benchmarks evaluating software construction and legal document handling.

To improve stability during multi-step workflows, Meta introduced two novel capabilities called Active Memory and Active Focus. These features are engineered to keep AI agents anchored to specified goals, reducing drift and preventing compounded errors when conducting broad research or generating complex code bases. Meta also emphasized safety training aimed at refusing prompts related to hacking or bioweapons creation, following earlier delays that Chief Executive Mark Zuckerberg attributed on X to safety refinements.

The launch marks a turnaround for Meta's AI division, which was restructured last year into Meta Superintelligence Labs under former Scale AI chief executive Alexandr Wang. After pioneering open-weight models with its Llama series, Meta temporarily lagged behind competitors like Google, Anthropic, and OpenAI. The lab previously released Muse Spark in April alongside a compact open-weight model named Muse Glimmer, though Meta has not shared technical details regarding the training or construction of the Muse Wonder family.

For enterprise users and developers, the arrival of Muse Wonder provides a competitive option for building agentic tools and automated programming assistants. In a company blog post, Meta called the releases a "major leap toward personal superintelligence," noting that the underlying technology will also power its personal agent, Muse, across consumer products.