Key Takeaways
- Meta’s new Muse Wonder Full ranks just below Claude Opus 5.5 and above Claude Fable 5.1 on the Artificial Analysis Intelligence Index, according to Meta’s own benchmarking.
- The models ship with Active Memory and Active Focus, features Meta says keep agents locked onto a stated goal instead of drifting into errors.
- Muse Wonder is available via an API and through Muse Code, and will power Meta’s Muse personal agent.
- Meta did not disclose training details or whether Muse Wonder-powered agents misbehaved during testing.
What Happened
Meta released two new AI models, Muse Wonder and Muse Wonder Full, which it says mark “the next step on our scaling ladder and a major leap toward personal superintelligence.” They follow Muse Spark, introduced in April, and the open-weight Muse Glimmer. Wonder is available through an API and through Muse Code, a tool for AI-assisted coding, and will power Meta’s Muse personal agent, which passed 500,000 users within a week of its own launch.
Why It Matters
Meta fell behind rivals after disappointing Llama releases while Anthropic, OpenAI, and Google advanced. CEO Mark Zuckerberg responded by funding Meta Superintelligence Labs and hiring former Scale AI chief Alexandr Wang to lead it. Meta says Muse Wonder Full now ranks just below Claude Opus 5.5 and above Claude Fable 5.1 on the Artificial Analysis Intelligence Index, and claims it beats GPT-6 Astra and both Claude models on specialized legal and coding benchmarks — claims not yet independently verified.
Technical Details
Muse Wonder adds two features, Active Memory and Active Focus, meant to keep an agent locked onto a stated goal during long tasks like research or large software projects, where drift compounds into costly mistakes. Meta also says it trained the model to refuse dangerous requests, including bioweapon instructions and hacking help, and that it scores well on resistance benchmarks. It did not say whether Muse Wonder agents misbehaved during its own testing.
Who’s Affected
Developers building on Meta’s API get a new frontier-tier option alongside Anthropic, OpenAI, and Google models, and Meta’s consumer products, including Muse, stand to inherit the upgrade. Zuckerberg said on X last month that Meta delayed the release specifically to do more safety work — a signal aimed at enterprise buyers and regulators watching the agent-safety debate.
What’s Next
Meta plans to release larger open-weight models beyond Glimmer and published a separate framework this month on scaling capability safely. Outside benchmark operators have yet to independently confirm Meta’s comparisons against Claude Opus 5.5 and GPT-6 Astra.
Source: Wired, Oct. 12, 2026.
