Meta unveils Muse Wonder and Muse Wonder Full, its new frontier AI models

A new model series aimed at the frontier

Meta on Monday, 12 October, unveiled Muse Wonder and Muse Wonder Full, two flagship artificial intelligence models that it says compete with the best systems from Anthropic and OpenAI. The launch is an attempt to catch up in the AI race as Meta tries to win back ground it lost in recent years. Meta says the models were previously known by the code-name Watermelon, and that they scored nearly as well as Anthropic's most advanced models on tests measuring reasoning, writing and coding, according to data it provided. Meta has not disclosed details of how the models were built or trained, beyond saying they were trained using its gigawatt-scale Prometheus data center.

How the models compare

Meta says the largest Muse Wonder version outperforms Anthropic's second-best model, Claude Fable 5.1, and ranks just below its best, Claude Opus 5.5, on the Artificial Analysis Intelligence Index, a widely used aggregated score of model capabilities. On some specialized benchmarks, including tests of legal document work and software building, Meta says the model also beats OpenAI's GPT-6 Astra and both Claude models. It is unclear how the models stack up against Google's Gemini 4 Argon, whose capabilities are still coming into focus. Alexandr Wang, who leads Meta Superintelligence Labs, told Axios that the models are competitive with the best in the world.

These are true frontier models, competitive with the best models in the world, and are particularly state of the art on personal and professional agentic work.

Memory, focus and access

Meta is emphasizing two capabilities, Active Memory and Active Focus, which it designs to help AI systems perform more useful work over extended periods. Active Memory keeps important information about a user across conversations and retrieves it when relevant to a new task, while Active Focus keeps an agent working toward a goal, including by trying different approaches when it hits obstacles. Muse Wonder is available through the Meta Model API and Muse Code, a tool for writing software with AI, while Muse Wonder Full is available to Muse Code subscribers. Meta says it plans to power its Muse agent with Wonder, starting with subscribers, but has not shared specifics or a timeline. In a blog post, the company said:

Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.

Background and Muse's rollout

Zuckerberg overhauled the company's AI research in 2025, sinking billions into Meta Superintelligence Labs and hiring former Scale AI CEO Alexandr Wang to lead it. Meta's Llama models were among the first frontier open-weight offerings from a big lab, but the company fell behind after several disappointing versions as Anthropic, OpenAI and Google rapidly introduced more capable models. The first model from the new lab, Muse Spark, was announced in April 2026, and Meta has also released Muse Glimmer, a small open-weight model, with larger open-weight models planned. The Muse agent app, which can operate software and apps on behalf of users, climbed to the top of the App Store in the weeks after its release in September. Meta says it has 3.5 billion users across Instagram, Facebook and WhatsApp.