Meta Launches Muse Wonder and Wonder Full Frontier AI Models to Power Personal Agents

UPDATE - follow-up to What is the 'Muse Wonder' model, and can it help Meta stay ahead in AI race?: Meta has unveiled its new Muse Wonder and Muse Wonder Full frontier AI models on Monday, October 12, designed to power personal and professional AI agents.

What Happened

Initially code-named 'Watermelon', the models were developed by the Meta Superintelligence Team led by Scale AI co-founder Alexandr Wang. They offer strong performance in coding and handling complex legal or financial questions. A key feature is 'Active Memory', which retains user context for weeks or months. Another is 'Active Focus', enabling agents to stay on task and persist until completion. The models can process text, image, video, and voice inputs, but only output text, with a context window of one million tokens at inference.

Performance and Access

On the Intelligence Index, the Wonder Full model scored 56, tying with Anthropic's Claude Sonnet 5.5 and outperforming OpenAI's GPT-6 Astra, Google's Gemini 4 Argon, and Anthropic's Claude Fable 5.1. The standard Muse Wonder model scored 52. On the SWE-Marathon software engineering benchmark, Wonder Full scored 57.5, beating Opus 5.5 and GPT-6 Astra. Developers can access Muse Wonder via the Meta Model API at $4 per million input tokens and $20 per million output tokens, as well as via Muse Code. Muse Wonder Full is available only for Muse Code subscribers.

Context

These models follow a turnaround for Meta after a prior slump in its AI efforts, prompting CEO Mark Zuckerberg to invest heavily in infrastructure and talent over the past 14 months. Meta aims to leverage its massive distribution network across apps like Instagram and WhatsApp to embed these agents deeply into everyday consumer lives.