Meta’s Muse Wonder returns it to the AI model race

A rebuilt research operation

Last year, Mark Zuckerberg invested billions of dollars in the newly formed Meta Superintelligence Labs and brought in former Scale AI CEO Alexandr Wang to lead it. The rebuild was difficult, and some researchers did not stay long.

Meta had once been an AI pioneer. Its Llama models were the first open frontier models from a major lab, and researchers and startups could download them and run them on their own hardware. Later releases disappointed users as Anthropic, OpenAI and Google moved quickly with more capable systems.

In April, Meta introduced Muse Spark, the new lab’s first model, alongside the smaller open model Muse Glimmer. The company said it planned to release larger open models in the future. It has not disclosed how Muse Wonder was built or trained.

What the benchmarks show

Meta says the largest version of Muse Wonder ranks above Anthropic’s Claude Fable 5.1 and below only Claude Opus 5.5 on the Artificial Analysis Intelligence Index, a composite measure of model capabilities.

On some specialist benchmarks, Meta says Muse Wonder also outperforms OpenAI’s top model, GPT-6 Astra, as well as Claude Fable 5.1 and Claude Opus 5.5. Those tests include work with legal documents and software development. Benchmarks are imperfect predictors of performance outside testing, but the results mark a substantial step forward for Meta’s models.

Memory, focus and the unanswered safety question

Muse Wonder adds two features intended to help AI agents stay on task: Active Memory and Active Focus. Meta says they are designed to keep a model aligned with its goal and prevent it from losing track of a task—an important concern in complex research and large software projects, where mistakes can be costly and compound.

Meta also says it trained the model to refuse dangerous requests, including instructions related to biological weapons or breaking into systems, and that it performed well on several benchmarks measuring resistance to such requests.

But recent cases involving AI agents from Anthropic, OpenAI and other companies show how easily powerful systems can stray when given vague instructions. Meta has not said whether it saw that behavior in agents powered by Muse Wonder during testing, or what it did specifically to address it. Last month, Zuckerberg wrote on X that Meta had delayed Muse models to work on safety. Earlier this month, the company described a framework in a blog post intended to make increases in model capability safer.

I think the benchmark results establish that Meta is back in the competition, not that it has solved the harder problem. The announcement gives us a ranking and new agent features, but little detail on how those features perform when instructions are ambiguous and errors accumulate. That gap matters more as the company pushes toward more capable systems.