Meta released two new AI models on Monday, Muse Wonder and Muse Wonder Full, TechCrunch and CNBC reported. TechCrunch called them the company's first attempt at a frontier model, the most advanced kind of AI system.

The models are larger than Meta's earlier Muse Spark model. Meta's own tests show Wonder and Wonder Full score about the same as rival top models, TechCrunch reported.

Meta says Muse Wonder is built to power digital assistants that do real tasks, like booking a dinner table. "We specifically tried to train the model to be able to try other approaches instead of stalling," Meta AI chief Alexandr Wang told CNBC. Meta also stressed what it calls "active memory," the model's ability to remember what users tell it over time, TechCrunch reported.

The model is not free. Users of the Muse agent app will need a $20 or $100 monthly plan to use it, CNBC reported. Developers can use Wonder through the Meta Model API for $4 per million input tokens and $20 per million output tokens, TechCrunch reported. Wonder Full will only be available to subscribers of Muse Code, Meta's coding agent.

Some users have complained online that the Muse agent read personal messages without permission, which Meta has disputed, CNBC reported. Wang said Meta is teaching users how the app works so that "consumers can trust the product going forward."

The models were trained at Meta's Prometheus data center in Ohio, TechCrunch reported.

Wonder Full is part of a Muse Code plan that costs $50 or $200 a month, CNBC reported. Muse rose to the top of Apple's App Store after Meta launched it last month.

What we know: Meta's Muse Wonder models are out now, mostly for paying users and developers.
What we don't: How the models perform in everyday use outside Meta's own tests.