Meta Reveals Two New Super Agents Capable of More Than You Think

Meta has launched its flagship models Muse Wonder and Wonder Full, designed to power autonomous agents in legal, programming, and business tasks. The models combine continuous memory and self-correction capabilities, outperforming GPT-6 Astra and Claude in global benchmarks. Alongside these achievements, the company is shifting to a payment model, locking features behind a paywall, with future integration into its apps.

Meta has launched its flagship models Muse Wonder and Muse Wonder Full, developed at the Meta Superintelligence Labs under the codename "Watermelon." The models are designed to power autonomous agents capable of acting on behalf of users in the real world, including managing insurance claims, deciphering rental contracts, building investment portfolios, analyzing medical data, and writing full code. The system integrates "active memory" and "active focus" mechanisms that allow for maintaining personal contexts and self-correcting errors. In performance tests, Wonder Full surpassed GPT-6 Astra and Claude Fable 5.1 in the Harvey LAB-AA legal agent test (13.1 vs. 8.6 and 6.4), the DeepSWE 1.1 programming test (79 vs. 74.1 for OpenAI), and the AutomationBench-AA business automation test (45.4 points). However, Meta's AI department head, Alexander Wong, admitted the model was launched with initial bugs. Meta is transitioning to a payment model: Wonder is available via API at $4 per million input tokens and $20 per million output tokens, while Wonder Full is exclusive to Muse Code subscribers at $50-200 per month. The company plans to integrate the agents into WhatsApp, Instagram, and Facebook, reaching 3.5 billion users. Capital expenditures are expected to surge to $145 billion this year.