Meta’s latest AI push has been years in the making: after rebuilding its research organization and pouring money into talent and computing, the company set out to turn its consumer reach into a credible challenge to OpenAI, Anthropic and Google.

In July, Meta AI chief Alexandr Wang reportedly told staff that the internal project then known as Watermelon had caught up to OpenAI, saying it used “an order of magnitude more compute” than the earlier Muse Spark family.¹ By Monday, that project had a public name: Muse Wonder and its larger sibling, Muse Wonder Full.

Mark Zuckerberg called Wonder “a much larger and more advanced class of models that brings us to the frontier,” framing the launch as the result of just over a year of work by Meta Superintelligence Labs.Mark Zuckerberg (finkd)² Meta says the models excel at personal and professional agents, aided by “Active Memory” that retains relevant information across conversations and “Active Focus” meant to keep an agent working through setbacks.³

There is some outside validation for that ambition. Aravind Srinivas, the Perplexity chief executive, congratulated Meta for training a “frontier-grade model,” saying Watermelon compared favorably with Fable 5.1 on WANDR evaluations.Aravind Srinivas (AravSrinivas)⁴ Artificial Analysis likewise placed Wonder Full at 56 on its Intelligence Index—above OpenAI’s GPT-6 Astra max, though behind Anthropic’s top Claude configuration at 58.⁵

But Meta’s frontier claim comes with caveats. The company’s own benchmarks remain to be broadly independently tested, and its most capable tools are not being handed to everyone. Muse users must pay $20 or $100 monthly for Wonder access; Wonder Full is reserved for Muse Code subscriptions costing $50 or $200 a month.⁶

That paid rollout also sharpens the stakes around trust. Meta says it has strengthened protections against cyber misuse and personal-data risks, yet users have complained that Muse accessed private messages without authorization—an allegation the company disputes. Wang’s response is that some users may have previously granted permissions and forgotten, while Meta works to make those choices clearer.⁶

The launch, then, is more than a benchmark contest. Meta has made its claim to the frontier; now it must show that agents with memory, autonomy and a subscription price can deliver without becoming a privacy liability.