Meta's Muse Wonder and Wonder Full, the first Superintelligence Labs models, claim frontier parity

Meta has shipped Muse Wonder and Muse Wonder Full, the first models from Meta Superintelligence Labs. The unit was formed about 14 months ago under Alexandr Wang after the Llama line was widely seen as falling behind the frontier. Meta positions the models for multistep reasoning, coding and autonomous agent workflows. It claims benchmark results comparable to Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra. The API-accessible Wonder is priced at $4 per million input tokens and $20 per million output tokens. The stronger Wonder Full sits behind a subscription.

The launch completes a tiered Muse lineup. Meta released the open-weight Muse Spark Open (816B total, 42B active) on October 16, tuned with Apple for dual Mac Studio inference. It also published a technical report on Muse Glimmer, a 30B open-weight agentic model built for a single consumer GPU at 4-bit precision. The infrastructure behind all of this is Prometheus, Meta's gigawatt-scale supercluster, detailed the same day.

Analysts at Investing.com framed the release as Meta's clearest break yet from its reputation for lagging. Matching closed rivals while also flooding the open-weight market gives Meta a dual position that neither OpenAI nor Anthropic occupies. The $4/$20 price undercuts typical frontier pricing, though not by a dramatic margin.

Early developer reaction is mixed. Commenters noted that independent benchmarks put API Wonder around 52 against Opus 5.5's 58. They described it as better than Astra for practical coding but weaker on hard terminal tasks, and some asked whether it is over-optimized for leaderboards. Early API rate-limit errors were also reported. Watch independent agentic evals and whether Wonder Full ever reaches the API.