Meta releases Muse Wonder, first frontier-class model from Alexandr Wang's lab

Meta's superintelligence lab, led by Alexandr Wang, shipped Muse Wonder, its first model the company positions as frontier-class. It follows Muse Spark, released in April, and arrives the same day as Anthropic's Fable 5.2. Wang claims state-of-the-art results on his team's internal agent-task evaluations, and on the independent Artificial Analysis Intelligence Index the model sits roughly level with top OpenAI and Anthropic systems.

The sources give little architectural detail, but the framing is agent-first: Meta is measuring Wonder on multi-step task completion rather than chat quality alone. That matches the rest of the Muse family, including the small Muse Ember model aimed at cheap tool calls and routing, released the same day.

Competitively, this is Meta's attempt to close the gap that opened after its Llama strategy stalled. Menlo Ventures' 2026 data shows Llama at just 7% of corporate open-source AI spend while Chinese models take 89%, and Muse is a closed family, a clear break from Meta's open-weights era. Reports also claim Google may accelerate Gemini 4 in response to Muse, though no official date exists.

Caveats: "roughly level" on a single index is not leadership, and Wang's headline claim relies on his own team's measurements. Meta is also cutting 8,000 jobs while lifting AI capex to $135-145B, so the pressure to show returns from the lab is high. Watch for independent agent benchmarks and API pricing.