$4 per million input tokens and $20 per million output tokens. That is the price Meta put on Muse Wonder, the first frontier model from its Superintelligence Lab, released on 12 October. TechCrunch noted it is roughly level with GPT-6 Sol or Opus 5.5.
Meta released two models, Wonder and Wonder Full. Wonder is available through the Meta Model API, while Wonder Full is limited to subscribers of Muse Code and is not served through the API. Both will power the Muse agent. No pricing was given for Wonder Full.
Meta’s model card says both reach benchmark scores comparable to Fable 5.1 and GPT-6 Astra, and that Wonder beats all three of Fable, Opus and Astra on MemoryBench.
The company is leaning on what it calls active memory, a model’s ability to retain and use what matters to a user over time.
Research scientist Kavya Kopparapu said the team put a tremendous amount of effort into making sure the model remembers what is important to the user.
It is a notable step for a company that has spent a year shipping lighter releases. Muse Spark, built for low latency, came out in April, and Meta had avoided competing head to head with OpenAI and Anthropic since Alexandr Wang reorganised its AI work.
The models were trained at Meta’s Prometheus data centre in central Ohio, which came online earlier this year.
The caveat is the evidence. Every comparison comes from Meta’s own testing, there are no independent results yet, and the company gave no parameter counts, training data or compute figures. Whether the weights will be released is also unstated.
