Meta releases Muse Wonder and Wonder Full frontier AI models

Meta has released Muse Wonder and Muse Wonder Full as its most capable models to date, designed for personal and professional agentic tasks. The company's Superintelligence Lab launched the models on Monday. Meta is introducing Muse Wonder and Wonder Full, our most capable models to date, built to deliver state-of-the-art performance on personal and professional agentic tasks, from navigating an insurance claim to making sense of a lease to building a working app, the company announced.

Models feature advanced memory and focus capabilities

The Wonder models include Active Memory and Active Focus features enabling sustained task performance and information retention across conversations. Active Memory allows the AI to retrieve information from across conversations so users do not have to start from scratch. Active Focus ensures the AI continues working on a task despite errors and will also prompt users for clarification or permission to perform tasks.

Meta emphasized the models' ability to retain and synthesize user information over time. Research scientist Kavya Kopparapu said, "We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond." The company highlighted its performance on the MemoryBench benchmark for continual learning, where Muse Wonder scores higher than models like Fable, Opus, or Astra.

Performance and availability details

Independent benchmarks show Muse Wonder places Meta at the AI frontier. On the Artificial Analysis Intelligence Index leaderboard, which uses a weighted composite score, Muse Wonder Full is second only to Claude Opus 5.5 Max. The same benchmarks show Muse Wonder leapfrogging GPT-Astra from OpenAI, Fable 5 from Anthropic, and Gemini 4 from Google.

Muse Wonder is available now to developers through the Meta Model API and Muse Code. Wonder Full is only available to Muse Code subscribers and is not served through the API. Meta said the hugely popular AI agent Muse surpassed 5 million app downloads and that Muse Wonder will soon power experiences within it, rolling out first to subscribers.

Context within Meta's AI strategy and competition

The release marks Meta's first frontier model launch since restructuring its AI efforts under Alexandr Wang, aiming to close the gap with OpenAI, Anthropic, and Google. Scale AI founder Alexandr Wang, then 28, was named Meta's new Chief AI officer in August of last year after Meta invested $14 billion to acquire and hire him. Since that restructuring, the company had released the lightweight Muse Spark model and the Muse agent but avoided state-of-the-art model releases.

The new Wonder models are larger and more capable than Muse Spark, with higher benchmark scores and better results on complex tasks like coding. Meta claims they can compete with Anthropic's Fable 5.1, OpenAI's GPT-6 Astra, and Google's Gemini 4 Argon in certain benchmarks. According to Meta’s testing, Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra. The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year.

Safety, pricing, and future integration

Meta pledged safety evaluations and privacy safeguards for the new models. The company said a full safety evaluation report will be available soon and that Muse Wonder models will avoid sharing private information without permission. Pricing for Wonder via the Meta Model API is set at $4 per million input tokens and $20 per million output tokens, which Meta says is roughly equivalent to GPT-6 Sol or Opus 5.5.

Both Wonder models will power the Muse agent. The launch follows the debut of the Muse AI agent early last month and comes six weeks after the release of Muse Spark 1.3, which debuted in April. Meta also unveiled the Muse Charm device, which will hit the market later this year. The company, which has invested tens of billions of dollars in AI, will need to demonstrate if its software successes translate to bottom-line results.