Meta launches Muse Wonder AI models, outperforming GPT-6 and Opus 5.5 in key benchmarks

Meta has announced Muse Wonder and Muse Wonder Full, a new family of frontier AI models designed to compete with OpenAI, Anthropic, and Google. Following the smaller Muse Spark 1.3 released last month, the new models were trained at Meta’s Prometheus data center using advances in pretraining and reinforcement learning. According to Meta’s benchmarks, Muse Wonder Full outperformed GPT-6 Astra and Claude Opus 5.5 on several coding and agentic evaluations, scoring 57.5% on SWE-Marathon and 79% on DeepSWE v1.1. It also led on the Harvey LAB-AA legal benchmark, although its Artificial Analysis Intelligence Index score of 56 remained slightly behind Opus 5.5 at 58.

Beyond performance improvements, the Wonder models introduce Active Memory, which retains relevant information across conversations, and Active Focus, which helps agents work through complex tasks, explore alternative approaches after failures, and request clarification when needed. Both capabilities are intended to support Meta’s Muse personal AI agent. The models also feature a 1-million-token context window and support text, image, audio, and video inputs, although outputs are currently limited to text.

Muse Wonder is available through the Meta Model API and Muse Code, while the flagship Muse Wonder Full is restricted to Muse Code subscribers. API pricing starts at $4 per million input tokens and $20 per million output tokens