Meta releases its first frontier model, Muse Wonder

Meta released Muse Wonder and Muse Wonder Full, which it presents as its first frontier-class models since Alexandr Wang restructured the company’s AI work. Where Muse Spark chased low latency in a lighter package, Wonder is the larger, more capable rung: Meta’s model card claims benchmark parity in the neighborhood of Fable 5.1 and GPT-6 Astra, with stronger results on complex tasks like coding.

API pricing for Wonder is $4 per million input tokens and $20 per million output — roughly aligned with GPT-6 Sol or Opus 5.5, per TechCrunch’s comparison. Wonder Full is gated to Muse Code subscribers and will not be served on the public API. Both models feed the Muse agent. Meta is leaning hard on “active memory”: retaining and synthesizing what a user shares over days and weeks, with MemoryBench scores it says beat Fable, Opus, and Astra.

This is Meta planting a flag that the Superintelligence Lab rebuild is producing frontier weights, not just consumer packaging — trained at the Prometheus data center in Ohio.

We have been waiting for Meta to stop circling the frontier and put a flagship model on the board. Wonder is that move. After Muse Spark and the Muse agent, the company needed a model story that could stand next to OpenAI and Anthropic on capability, not only on distribution. Claiming comparable scores to Fable 5.1 and GPT-6 Astra, plus a MemoryBench lead, is how you reopen that conversation.

The product shape is telling. Wonder on the Meta Model API at frontier-ish prices says Meta wants developers in the tent. Locking Wonder Full behind Muse Code, and routing both into the Muse agent, says the real customer is still Meta’s own assistant stack — personal superintelligence as Zuckerberg has been branding it. Active memory is the thesis: an agent that remembers what matters across long stretches is more valuable than a slightly higher coding bench if you live inside Meta’s apps.

Credit where it is due. Shipping a frontier pair from Prometheus after a rocky lab rebuild is operationally hard. Pricing in the same band as peer models signals confidence rather than a dumping war. We want open, competitive model markets; another serious lab at the table is good for builders.

Eyes open on the rest. Benchmarks are Meta’s own presentation until outsiders reproduce them. Wonder Full’s API absence concentrates the best weights inside Meta’s subscription surface. Long memory is a gift for useful agents and a liability for privacy and mistaken retention — especially for a company whose trust problems are not theoretical. And “major leap toward personal superintelligence” is launch language; the durable test is whether Muse users feel the model stay coherent for weeks without drifting, leaking, or over-remembering.

Our read: Meta is back in the frontier race on paper. The next proof is third-party evals, API reliability at that price, and whether active memory behaves like a careful companion or a clingy profile engine. I’m watching independent Intelligence Index and coding-bench checks, Muse Code conversion around Wonder Full, and how Meta documents what the model keeps and forgets.

Meta’s Superintelligence Lab, led after last year’s overhaul by Alexandr Wang, had released Muse Spark and the Muse agent but had not yet fielded a state-of-the-art frontier model in the style of OpenAI and Anthropic. Muse Wonder is framed as the next step on Meta’s scaling ladder, trained at Prometheus in central Ohio.