Muse Wonder Full outperformed GPT-6 Astra, but not Claude Opus 5.5.

Muse Wonder Full scored higher than GPT-6 Astra and Gemini 4 Argon in the independent Artificial Analysis index. But it did not take first place: Claude Opus 5.5 retained the lead. For users, this is a useful caveat to the attention-grabbing headline, because the result depends on which task is being evaluated.

According to Artificial Analysis, Wonder Full scored 55.8 points, GPT-6 Astra 52.7, Gemini 4 Argon 52.6, and Claude Opus 5.5 57.6. Meta separately claims strong results in agentic and coding tests. For example, in SWE-Marathon, Wonder Full scored 57.5% versus 46.3% for Opus 5.5.

Meta’s main focus is agent behavior. Active Memory is supposed to carry important information about the user across conversations. Active Focus teaches the model to continue working on a complex task, try different approaches, and clarify details before taking important actions. If these features work as the company describes, their practical value will be in tasks that take more than one conversation.

There is also an inconvenient part for the bold claims. Wonder Full lost to Opus 5.5 in ProgramBench and Terminal-Bench, and trails it in the overall Artificial Analysis index. In the AA-Omniscience test, Wonder Full scored 30 points versus 46 for Opus 5.5, meaning that strong agentic results do not imply equally strong knowledge. The model weights have not been published, and Wonder Full is available only to Muse Code subscribers.

For which task would you be more willing to trust an agent to work without constant supervision: programming or personal matters where remembering your preferences is important? Source: All articles in a row / Machine learning / Habr