Meta has introduced Wonder and Wonder Full, advanced AI systems designed to rival the capabilities of OpenAI and Anthropic. These frontier models will be available to paying subscribers through the Muse platform.
The shift from Llama to rival ChatGPT and Claude
Meta is fundamentally pivoting its AI strategy with the release of Wonder. Unlike the previous Llama lineage, which focused on open-source accessibility, the Wonder models are proprietary and aimed at the highest tier of performance. According to the report, independent benchmarking conducted by Artificial Analysis indicates that Wonder outperforms previous Meta systems as well as select models from Google, OpenAI, and Anthropic.
However, this leap in performance comes with a significant financial burden. The report notes that while Wonder ranks among the market's most capable systems, it is also among the most expensive to operate . This suggests that Meta is prioritizing raw power and competitive parity with ChatGPT and Claude over the operational efficiency seen in its smaller, open-source iterations.
Alexandr Wang and the $14 billion Scale AI catalyst
The development of the Wonder and Wonder Full models was steered by Chief AI Officer Alexandr Wang. Mark Zuckerberg recruited Wang to lead the Superintelligence Labs division following a massive $14 billion investment in the San Francisco-based startup Scale AI in June 2025. This recruitment signals a shift in Meta's internal leadership,moving toward a more aggressive, "frontier" approach to AI developent.
Under Wang's guidance, Meta is attempting to transition from providing simple conversational interfaces to creating truly autonomous agents. These agents are intended to move beyond text generation to execute complex, real-world tasks, marking a new chapter in Meta's pursuit of artificial general intelligence.
From Muse Charm hardware to a $145 billion capital spend
Meta is integrating these new models into a comprehensive ecosystem of "agentic" software and hardware. This includes the launch of Muse Code and a handheld hardware device known as Muse Charm, which is designed to handle research, bookings, and coding tasks. As reported, Meta indicated in July that its annual AI capital expenditure could reach $145 billion this year to support this infrastructure.
This aggressive spending is an attempt to leverage Meta's massive reach—which includes more than 3.6 billion daily active users across Facebook and Instagram—to diversify its income. By offering paid subscriptions for the Muse platform, Meta is seeking to reduce its historical reliance on advertising revenue.
The September White House commitment and the risk of escaped models
Despite the technological push, safety remains a critical friction point. In September, Mark Zuckerberg joined executives from Nvidia, xAI, Google, OpenAI, and Anthropic to sign a White House commitment to implemment rigorous safety frameworks. Meta has stated that the release of the Muse platform was intentionally delayed to strengthen these safeguards.
However, the report highlights a climate of caution, noting that competitors have warned of experimental models that took unauthorized actions or escaped secure environments. It remains unclear which specific models were involved in these breaches or what the exact nature of the "unauthorized actions" entailed, as the source does not provide specific examples or name the companies whose models escaped.
Comments 0