Meta finally has a model it’s willing to put in the same sentence as the industry’s best. The company’s Superintelligence Lab has released Muse Wonder and Muse Wonder Full, its first serious attempt at a frontier AI model and the clearest sign yet of what the expensive reorganisation of its AI efforts was meant to produce.
The pair sit above Muse Spark, the lightweight, low-latency model Meta shipped earlier, and are pitched as larger systems built for harder work such as coding and complex multi-step tasks. According to Meta’s own model card, both deliver benchmark results comparable to Claude Fable 5.1 and OpenAI’s GPT-6 Astra. Those are self-reported numbers, and independent testing will matter far more than launch-day charts, especially for a company whose Llama 4 release last year drew criticism over how its benchmark results were presented.
Meta’s real differentiator is memory. The company is leaning heavily on what it calls active memory, the ability to hold onto and connect information a user shares across days, weeks and months. On MemoryBench, a research benchmark designed to test continual learning and memory construction, Meta says Muse Wonder outperforms Fable, Opus and Astra. That focus fits Meta’s stated goal of building long-running personal assistants, and it plays to an obvious strength. Few companies know more about their users’ lives than the one running Facebook, Instagram and WhatsApp, which also means the privacy questions around a model designed to remember everything will not go away quietly.
Pricing confirms Meta is competing at the top end. Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens, putting it in the same bracket as GPT-6 Sol and Claude Opus 5.5. The more powerful Wonder Full won’t be offered through the API at all and is reserved for Muse Code subscribers, while both models will power the Muse agent. That’s a sharp shift from the Llama era, when Meta’s AI identity was built around freely available open-weight models.
This is also the first frontier release since Alexandr Wang, founder of Scale AI, took over and restructured Meta’s AI organisation in August 2025. Until now, the new setup had produced smaller models and an agent product rather than anything aimed directly at OpenAI or Anthropic. The models were trained at Prometheus, Meta’s massive Ohio data center that came online this year.
Meta has spent enormous sums on talent and compute to reach this point. Matching rivals on benchmarks is the entry ticket, not the win. The harder test is whether developers and everyday users actually choose Muse over models they already rely on.
