Meta unveils Muse Wonder, says it matches top AI rivals

Meta Platforms on Monday released Muse Wonder and Muse Wonder Full. The company calls them its most advanced artificial intelligence models and says they compete with the best systems from OpenAI and Anthropic. Bloomberg reported that the release is the company's most capable model since Chief Executive Mark Zuckerberg rebuilt his AI operation through an aggressive hiring push, an effort it described as a 14-month campaign. The launch tests whether that multibillion-dollar spending has paid off.investing+2

Muse Wonder is available to developers through the Meta Model API and Muse Code, Meta's AI-assisted programming tool. The more capable Muse Wonder Full is limited to Muse Code subscribers. According to Axios, Muse Wonder will also power future features in the Muse assistant, at first only for paid subscribers. Meta shares rose after the announcement.axios+1

Claims of frontier performance

Meta posted benchmarks showing the models, code-named "Watermelon," beating or matching the latest offerings from OpenAI and Anthropic. Wired reported that, on the Artificial Analysis Intelligence Index, Meta says the largest variant ranks just behind Anthropic's top model, Claude Opus 5.5. The company also says Muse Wonder beats OpenAI's GPT-6 Astra on certain specialized tests, including legal-document and software-development benchmarks.wired+1

"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," Alexandr Wang, head of Meta Superintelligence Labs, told Axios.axios

The models introduce two features. "Active Memory" retains information about a user across conversations. "Active Focus" keeps an agent working toward a goal by trying different approaches, and it can ask for clarification when needed. Meta says the models were trained at its gigawatt-scale Prometheus data center.investing+1

Safety and spending

Meta says it trained the models not to share private information without permission and to refuse requests that could cause harm. It also says it built them under its Meta Superintelligence Scaling Framework. A model card was due Monday, with a fuller report to follow. Wired noted that Meta has not said whether it saw any misbehavior from Muse Wonder agents during testing.wired+2

Zuckerberg told the Financial Times that Meta has caught up with its rivals at the frontier. He rejected calls to slow AI development and said safety should be the responsibility of individual labs. Asked about open-sourcing the Wonder family, Wang would not commit, but said, "We're still committed to open source".investing+1

The release follows Meta's Sept. 8 launch of its Muse personal agent. That agent can send emails and make payments on a user's behalf and is offered in a free tier and in subscriptions priced at \$20 and \$100 a month. The question now is whether independent testing confirms Meta's performance claims.axios+1