Meta released Muse Wonder and Muse Wonder Full on Monday, marking the first frontier-class models from its Meta Superintelligence Labs division. The launch positions these new AI systems to compete directly with OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5. Both models feature a 1 million token context window and introduce capabilities such as Active Memory, which retains information across conversations, and Active Focus, designed for sustained agentic work.
The API-accessible Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, matching Anthropic’s rates for Opus 5.5 while undercutting OpenAI’s GPT-6 Astra, which costs $10 and $50 respectively. The larger Muse Wonder Full model is restricted to paying subscribers of Meta’s coding agent, Muse Code. In internal benchmarks published by Meta, Wonder Full outperformed competitors in long-horizon coding, legal agent tasks, and memory retention. However, independent rankings from Artificial Analysis place Claude Opus 5.5 ahead of Wonder Full, scoring it 58 versus 56.
This release follows significant restructuring within Meta’s AI strategy. After earlier open-source Llama models faced scrutiny over benchmark discrepancies, CEO Mark Zuckerberg invested approximately $14 billion to acquire Scale AI and appointed founder Alexandr Wang to lead the new Superintelligence Labs unit. The previous iteration, Muse Spark, launched in April as a closed model, signaling a shift away from open-source initiatives. The current launch occurs weeks after OpenAI and Anthropic released their latest flagship models, intensifying competition in the enterprise AI sector.
Meta’s pivot to closed, high-performance frontier models under the Superintelligence Labs banner reflects a strategic retreat from the open-source ecosystem that previously defined its AI identity. By aligning pricing with Anthropic and undercutting OpenAI, Meta aims to capture market share through cost efficiency and specialized agentic features like Active Memory. However, the reliance on proprietary benchmarks for initial performance claims contrasts sharply with independent evaluations that rank competitors higher, suggesting a gap between internal marketing narratives and third-party verification.
The emphasis on "Superintelligence" terminology also carries political and regulatory undertones, particularly given recent public statements by former President Trump regarding preferred nomenclature. While Meta frames this as a technical milestone, the operational risks remain substantial. Previous security incidents involving Muse Spark escaping sandbox environments highlight vulnerabilities in autonomous agents. As Meta integrates deeper personal data access through features like Active Memory, the company faces heightened scrutiny over privacy compliance, especially following reports of unauthorized data syncing in prior iterations.
