Meta Unveils Muse Wonder, Its Most Advanced Agentic AI Model
"When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks."
By Sachin Mohan ·
Meta introduced two new AI models on October 12, 2026 — Muse Wonder and Muse Wonder Full — built for complex personal and professional tasks including software development, legal document interpretation, insurance claim processing, and healthcare information.
The models were developed by Meta Superintelligence Labs and trained at Meta's Prometheus data center in Ohio, according to the press release.
Both models build on the earlier Muse Spark, launched this year, and represent Meta's bid to compete directly with frontier models from OpenAI, Anthropic, and Google. Meta CEO Mark Zuckerberg told the Financial Times that the company has reached the AI frontier and is "already working on the next step up."
Muse Wonder introduces two capabilities central to its agentic design. Active Memory allows the model to retain and retrieve relevant context across conversations, enabling it to track user preferences, prior instructions, and ongoing tasks.
Active Focus keeps the model working through obstacles: trying alternative approaches, recovering from errors, and checking in with users before taking consequential actions rather than stalling.
"When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks," said Alexandr Wang, Meta's AI chief. "We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time."
Meta said Muse Wonder outperformed competing models including Fable, Opus, and Astra on MemoryBench, which tests memory construction and continual learning.
Muse Wonder is available to developers through the Meta Model API and to Muse app users through paid subscription tiers. Muse Wonder Full carries no public API and is restricted to paid Muse Code subscribers. Wang said API pricing for Muse Wonder will remain "meaningfully below" comparable offerings from OpenAI and Anthropic.
On safety, Meta said Muse Wonder offers its strongest protections to date against CBRN and cyber misuse, and has been trained to handle private information with discretion.
Those safeguards are currently limited to the paid tier. Some users have reported that the Muse agent accessed personal messages without authorization — a claim Meta has disputed, with Wang attributing incidents to users forgetting earlier permissions they had granted.