Meta launches its first frontier AI models
What's the story
Meta has launched its first frontier models, Muse Wonder and Muse Wonder Full. The new additions from the Superintelligence Lab are bigger and more capable than their predecessor, Muse Spark. They also promise better performance on complex tasks such as coding. According to Meta's tests, these models achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra.
Cost details
Muse Wonder priced similarly to GPT-6 sol
Muse Wonder can be accessed via the Meta Model API at $4 per million input tokens and $20 per million output tokens. This pricing is similar to that of GPT-6 Sol or Opus 5.5.
However, Muse Wonder Full will only be available to paid Muse Code subscribers with monthly plans between $50 and $200.
The advanced model offers more reasoning and processing power for complex engineering tasks but won't be available through the public API.
Tech advancement
Muse Wonder outperforms other models on MemoryBench benchmark
The new models come with a focus on retaining and synthesizing information provided by users, a feature Meta calls "active memory." This is part of the company's vision for personal superintelligence.
To improve recall, Muse Wonder has outperformed Fable, Opus, and Astra on the MemoryBench benchmark.
Research scientist Kavya Kopparapu said they worked hard to ensure that the model remembers what's important to users over days, weeks, months.
Model features
Muse Wonder can be used in app and game development
Muse Wonder is designed for more complex tasks such as software development, healthcare information, finance, and legal information.
The model can be used in app and game development as well as examining rental and insurance claims.
It also comes with a feature called "Active Focus," which helps the model recover from technical errors, explore alternative solutions, and ask users for clarification when needed.