Meta unveiled a pair of advanced artificial intelligence systems on Monday, dubbed Muse Wonder and Muse Wonder Full, as the tech company attempts to close the gap with industry leaders Anthropic and OpenAI. The rollout marks the company’s first release of a frontier-level system following billions of dollars in infrastructure investments, including training at its new Prometheus data facility in central Ohio.
Chief Executive Mark Zuckerberg declared on social media that the new line represents a significantly larger and more capable tier of technology that brings Meta to the industry frontier. The models expand upon Muse Spark, a lightweight architecture released earlier this year, and are designed to power both developer platforms and Meta’s consumer assistant app, Muse.
New memory and reasoning tools
According to Meta, the Wonder models are engineered to execute intricate, multi-step actions on behalf of users. The company highlighted capabilities such as drafting software applications, evaluating residential lease contracts and processing insurance claims. Two specific mechanisms underpin the new software: Active Memory and Active Focus.
The company said Active Memory lets the system retain and synthesize details shared across interactions over extended periods. Kavya Kopparapu, a research scientist at Meta, stated in an introductory video that the team dedicated substantial effort toward ensuring the model retains crucial user details across days, weeks and months. Meta reported that the software outperformed rival models, including Astra, Opus and Fable, on the MemoryBench evaluation framework.
Meanwhile, the Active Focus tool is intended to help the system navigate unexpected errors. Rather than halting when confronted with technical hurdles, the model can adjust its strategy, formulate alternative solutions and seek clarification from the user, CNET reported.
Commercial pricing and developer access
Meta is using the launch to test customer appetite for paid software tiers. The standard Muse Wonder model is accessible through the company’s application programming interface, with subscription tiers priced at $20 and $100 each month, alongside usage fees of $4 per million input tokens and $20 per million output tokens, TechCrunch reported.
The larger variant, Muse Wonder Full, provides expanded reasoning capacity and greater computing power for corporate workflows and engineering tasks. It will not be offered through a public interface, instead remaining exclusive to subscribers of Meta’s Muse Code engineering tool under monthly plans ranging from $50 to $200.
Alexandr Wang, the Scale AI founder who reorganized Meta’s artificial intelligence division in August of last year, said the model’s enhanced capabilities justify the higher fees compared to basic consumer subscriptions, though he maintained that rates remain lower than those of chief rivals.
Debate over safety and self-regulation
The introduction of the high-end software comes amid persistent debate across Silicon Valley regarding the speed of technological progress. Prominent industry leaders, such as OpenAI Chief Executive Sam Altman and Anthropic Chief Executive Dario Amodei, have previously urged caution and external regulation. Anthropic recently withdrew access to certain advanced models after federal authorities instituted strict usage restrictions, according to CNET.
Zuckerberg, by contrast, has consistently maintained that technology firms can manage risks internally without government-mandated pauses. In an interview with the Financial Times, Zuckerberg said he favors the view that artificial intelligence can be safely built using standard guardrails rather than viewing it as an uncontrollable force.
The rollout also follows privacy concerns among early users of the Muse personal assistant, which became the most downloaded program on Apple’s App Store following its debut last month, CNBC reported. Meta disputed complaints on social media alleging the program accessed private correspondence without consent, with Wang stating that some customers may have granted permissions early in conversations and later forgotten doing so.
It remains to be seen whether enterprise clients and software engineers will embrace Meta’s premium pricing or stick with established offerings from its competitors.
