Meta Ships Its First Frontier Model, Muse Wonder

Meta has released its first model at the "frontier" level, the most advanced tier that AI labs build. It came out of Superintelligence Lab, Meta's AI research division. The models are Muse Wonder and Muse Wonder Full, trained at Meta's Prometheus data center in Ohio. Access to Wonder costs $4 per million input tokens and $20 per million output tokens. Tokens are the chunks of text an AI model reads and writes. That price is roughly the same as OpenAI's GPT-6 Sol or Anthropic's Opus 5.5. Wonder Full is available only to subscribers of Muse Code, Meta's separate coding-assistant service. It is not offered through the API, the programming interface other software uses to connect to the model. By Meta's own tests, Wonder scores higher than Fable and Opus, both Anthropic models, and OpenAI's Astra. That is on MemoryBench, a benchmark for "active memory." Active memory is a model's ability to store and summarize what it learns about a user over weeks and months.

These MemoryBench results come only from Meta's own testing. No independent benchmark has confirmed them yet. Pricing it at the same level as Opus 5.5 means Meta is claiming parity with top models, not undercutting them on price. Both claims deserve some caution until independent tests back them up.