After months of trailing OpenAI and Anthropic, Meta says its larger Muse Wonder models put it among the leaders, with a consumer-focused strategy and a closed approach, though the launch was marred by early bugs.
Meta has launched Muse Wonder, a new family of AI models that chief executive Mark Zuckerberg says lifts the company to the cutting edge of the field after a period of playing catch-up with OpenAI and Anthropic.
Announcing the release on X, Zuckerberg said the earlier Muse Spark line was the company’s first family of models, while Muse Wonder is a considerably larger and more capable class. He told the Financial Times it would enable far more sophisticated behaviours than before. Meta said the model, known internally as Watermelon, outperforms rivals such as Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Astra on certain tasks, citing its own benchmarks and data from Artificial Analysis.
Zuckerberg acknowledged that reaching this point took heavy spending on computing power and long, rigorous scientific work. Meta has put billions of dollars into AI over the past year. Last June it invested $14 billion in Scale AI and brought in the firm’s co-founder, Alexandr Wang, who now heads Meta’s Superintelligence Labs. Wang described the launch as “the big moment” and said the models are competitive with the best available, particularly in personal and professional agentic work.
Built for consumers, and now closed
Two versions have been released: the standard Muse Wonder and the more powerful Muse Wonder Full. The standard model will power the paid tier of Muse, the personal AI assistant Meta introduced last month. Muse uses autonomous bots for tasks such as ordering groceries and booking travel, and Meta says the app has been the most downloaded on both Apple and Android stores in the US.
Wang said that while OpenAI and Anthropic have concentrated largely on coding agents, Meta chose a different path by building large models for broad consumer use. He said Muse Wonder has “active memory” that lets it retain relevant information over days, weeks and months, and keep working through dead ends by trying new approaches.
The models are available through the Meta Model API and Muse Code. Pricing starts at $0.20 (about Rs 19) per million tokens for basic use and rises to $200 (about Rs 19,000) a month for Muse Wonder Full. The rollout hit trouble, however, when a bug distorted how usage counted against limits for the Code subscription. Wang later said on X that several bugs were being investigated.
The release also marks a break from Meta’s earlier stress on open-source AI. Muse Wonder will be closed, with its best versions sold directly to businesses. Zuckerberg hinted at more to come, saying work was already under way on the next step up and that progress would likely accelerate.
Costs and scrutiny at home
The push comes alongside cost-cutting and legal pressure. In May, Meta laid off 10 per cent of its workforce, about 8,000 jobs. It also planned software to track employees’ mouse movements and keystrokes to train AI models, but withdrew the idea after staff backlash. In August, the company agreed to pay $18 billion to settle a case brought by US states over alleged failures on child protection and social media addiction.
