On October 12, Meta Superintelligence Labs (MSL) released Muse Wonder and Muse Wonder Full, the first Meta models the company calls “frontier-level”, marking Meta's formal entry into the frontier-model race.
Three things to note
1. Benchmarks put it among the frontier models. The Intelligence Index from independent evaluator Artificial Analysis gave Wonder Full 56 points, level with Claude Sonnet 5.5 and ahead of GPT-6 Astra, Gemini 4 Argon and Fable 5.1 (all at 53), behind only Claude Opus 5.5 (58). The standard Wonder scored 52. Meta's model card says the Wonder series performs on par with Fable 5.1 and GPT-6 Astra on complex tasks such as coding.
2. The focus is “Active Memory”. Unlike its lightweight, low-latency predecessor Muse Spark, the Wonder series is optimized mainly for long-term memory: it remembers important user information over days to months and carries context across multiple conversations. On Meta's in-house MemoryBench, it scored higher than Fable, Opus and Astra. Research scientist Kavya Kopparapu said the team “put a lot of effort into making the model remember what matters to you, across days, weeks, months and beyond”.
3. The Llama open-source era is over. Neither model has open weights, a complete break from the open approach of the Llama era. Even more striking, the top model, Wonder Full, is not available through an API at all and is offered only to Muse Code subscribers. The standard version is offered through the Meta Model API at $4 per million input tokens and $20 per million output tokens, comparable to GPT-6 Sol.
Both models were trained at Meta's gigawatt-scale Prometheus data center in Ohio and built over about 14 months by the Superintelligence Labs, led by former Scale AI CEO Alexandr Wang. They will be used to power the Muse agent. In a post on X, Mark Zuckerberg said Muse Wonder lets Meta “reach the frontier”, and he rejected calls to slow down development. Meta's shares rose more than 3% after the release.
Source: read the original. This text is a summary for reference only.
