Meta has launched Muse Wonder and Muse Wonder Full, a pair of artificial intelligence models that the company describes as possessing frontier capabilities, positioning them among the world's most powerful AI systems. The release, announced today in a company blog post, marks a significant shift for Meta, which had fallen behind competitors after earlier leading the industry with its Llama series of open-weight models. The new models are designed to power features across Meta's platforms, including its Muse personal agent, and will be accessible through an API and Muse Code, a software development tool.
According to Meta, the largest version of Muse Wonder surpasses Anthropic's second-tier model, Claude Fable 5.1, and sits just beneath its top-performing Claude Opus 5.5 on the Artificial Analysis Intelligence Index, an aggregated measure of model capabilities used widely across the industry. The company also reports that Muse Wonder outperforms OpenAI's most advanced model, GPT-6 Astra, as well as both Claude Fable 5.1 and Claude Opus 5.5 on specialized benchmarks that evaluate models' abilities to handle legal documents and build software. Meta released its first model from the newly formed Meta Superintelligence Labs, Muse Spark, in April this year, and has since introduced a smaller open-weight model called Muse Glimmer, with plans to release larger open-weight versions in the future.
The company states in its announcement that the new models represent "the next step on our scaling ladder and a major leap toward personal superintelligence." Meta says it has invested substantial resources in training Muse Wonder to refuse generating dangerous outputs, such as instructions for creating bioweapons or assistance with hacking, and reports strong performance on benchmarks designed to measure models' resistance to such prompts. The models include two new features, Active Memory and Active Focus, built to maintain the model's focus on stated objectives and prevent confusion during complex tasks like extensive research or large software projects, which can otherwise lead to costly errors or cascading mistakes.
Meta's resurgence comes after CEO Mark Zuckerberg revamped the company's AI research last year, investing billions in Meta Superintelligence Labs and recruiting former Scale AI CEO Alexandr Wang to lead the initiative. The effort has been turbulent, with some researchers departing quickly. Meta's earlier dominance in AI, built on its Llama series that could be downloaded and operated on personal hardware, eroded after the company delivered several underwhelming model versions while Anthropic, OpenAI, and Google rapidly advanced. Zuckerberg indicated in a post last month that Meta delayed the Muse models' release to address safety concerns, and the company published a framework earlier this month outlining how it plans to scale model capabilities safely. Though Meta hasn't disclosed whether Muse Wonder agents misbehaved during testing or what specific steps were taken to prevent such issues, recent incidents involving rogue AI agents from competitors have highlighted how easily powerful models can deviate when given open-ended instructions.
The launch positions Meta once again as a serious contender in the escalating race to develop the most capable AI systems, with Muse Wonder likely to intensify competition across the industry. Meta did not reveal details about how the new models were constructed or trained, and benchmarks don't always predict real-world performance reliably, but the advancement represents a substantial leap forward for the company's AI capabilities. For organizations deciding where to place their AI infrastructure bets, the sudden return of a major player scrambles assumptions about which platforms will dominate the agent economy. Meta's willingness to toggle between open and closed approaches adds another layer of uncertainty for enterprises trying to build on stable foundations.