Meta releases its first frontier model, Muse Wonder

Meta’s Superintelligence Lab put out two new models on Monday: Muse Wonder and Muse Wonder Full. That makes them the company’s first real swing at a frontier model, after Muse Spark had been positioned as the lighter, faster option. Wonder is the bigger bet. Meta says it performs better on harder work like coding and posts benchmark scores that stack up against Fable 5.1 and GPT-6 Astra.

The pricing makes the shift pretty obvious. Through the Meta Model API, Wonder will run $4.00 per million input tokens and $20.00 per million output tokens, which puts it in the same general cost bracket as GPT-6 Sol or Opus 5.5. Wonder Full is even more locked down: it will be available only to Muse Code subscribers and won’t be served through the API at all. Both models will also power the Muse agent.

Meta is also leaning hard into memory. The company says the new models are better at keeping and combining information users share over time, something it calls “active memory.” To make that case, Meta pointed to MemoryBench, a benchmark for continual learning and memory construction. On its own testing, Wonder scores higher there than Fable, Opus or Astra.

That focus fits Meta’s long-running push for assistants that can sit with a user for days, weeks, and months, not just answer a single prompt and vanish. Research scientist Kavya Kopparapu said the team spent a “tremendous amount of effort” on remembering what matters over longer stretches. The release is also a shift in cadence: these are Meta’s first frontier models since Alexandr Wang restructured its AI work in August of last year. Since then, the company had been shipping lighter pieces like Muse Spark and Muse agent instead of going toe to toe with OpenAI and Anthropic on flagship models.

The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year.