Meta Debuts Muse Wonder AI Models in Bid to Match Frontier Competitors

Meta has introduced a new family of artificial intelligence models called Muse Wonder, marking the company's entry into the top tier of frontier AI performance. The launch follows aggressive investments by Meta over the past year in technical talent and computing infrastructure to close capability gaps with rival research laboratories.

According to independent benchmark evaluations from Artificial Analysis, reported by Mashable Tech (https://mashable.com/tech/meta-launches-muse-wonder-ai-model-family), the high-tier Muse Wonder Full model placed second overall on the platform's Intelligence Index leaderboard. The score puts the model in a tie with Anthropic's Claude Sonnet 5.5 Max, trailing only Claude Opus 5.5 Max while surpassing competing frontier systems including OpenAI's GPT-Astra, Anthropic's Fable 5, and Google's recently launched Gemini 4.

The rollout comes six weeks after Meta released Muse Spark 1.3, an iteration of a model family that first debuted in April. Over the past year, Meta Chief Executive Mark Zuckerberg moved to expand computing capacity and recruit key personnel, culminating in a $14.3 billion acquisition of Scale AI and the appointment of its founder, Alexandr Wang, then 28, as Meta's Chief AI Officer.

Wang stated that the model family was developed under the internal codename "watermelon," according to reporting cited by Mashable Tech from CNBC. Meta designed Muse Wonder primarily for agentic workflows, focusing on multi-step consumer and enterprise tasks such as parsing lease contracts, navigating insurance claims, and writing functional code.

To support long-running agent workflows, Meta introduced two architectural capabilities: Active Focus and Active Memory. Active Focus is designed to allow an agent to maintain progress during complex tasks, test alternative solutions, recover from errors, and request user clarification before executing critical actions. Active Memory enables the system to persist context and preferences across separate sessions.

Meta also addressed safety and privacy controls following prior scrutiny over data handling in its consumer AI products. The company stated that Muse Wonder is configured to avoid unauthorized sharing of private user data, reject harmful prompts, and assist enterprise users with cybersecurity defense. Meta noted that a comprehensive safety evaluation report will be published in the near future.

Artificial Analysis reported that Muse Wonder Full placed in the top three across six of ten evaluations on its composite Intelligence Index, exhibiting particular strength in scientific reasoning and agentic knowledge work.

Meta has made Muse Wonder available to developers through the Meta Model API and Muse Code, while access to Muse Wonder Full is restricted to paid Muse Code subscribers. Meta plans to integrate the new models into its consumer AI agent Muse—which has surpassed 5 million app downloads—for subscribers paying $20 or $100 per month.