Meta Launches Muse Wonder AI Models to Challenge OpenAI

Meta introduced two advanced artificial intelligence models on Monday, October 12, 2026, as the company intensifies its competition with OpenAI and Anthropic. The models, Muse Wonder and Muse Wonder Full, are designed to improve digital assistants, strengthen coding capabilities and support Meta's efforts to generate subscription and business revenue from AI technology.

According to Meta's published testing results, the models achieved benchmark scores approaching those of leading competitors in reasoning, writing and coding. The company plans to integrate Muse Wonder into its Muse personal AI agent, while reserving Muse Wonder Full for subscribers to its AI coding service, Muse Code.

Meta said the models were developed under the internal codename Watermelon. Their release follows a restructuring of the company's AI division led by Alexandr Wang, who joined Meta as chief AI officer in June 2025 after founding AI startup Scale AI.

In an interview, Wang said the development effort focused on helping AI agents complete complicated tasks without stopping whenever they encounter obstacles.

"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," Wang said.

He described the training approach as an effort to make agents more adaptable when their initial methods fail, particularly when performing practical tasks on behalf of users.

"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." Wang said.

Meta is also emphasizing memory capabilities that allow its models to retain and organize information supplied by users over extended periods. According to TechCrunch, the company reported that Muse Wonder outperformed several competing models on MemoryBench, a benchmark designed to assess memory construction and continual learning.

Research scientist Kavya Kopparapu discussed the company's focus on retaining relevant personal information during a video accompanying the launch.

"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," Kopparapu said.

Pricing reflects Meta's effort to turn its AI investments into new sources of income beyond digital advertising. According to CNBC, access to Muse Wonder through the Muse app requires a monthly subscription costing $20 or $100, depending on the plan. The app itself remains free, with paid options offering additional usage.

TechCrunch reported that access through Meta's Model API costs $4 per million input tokens and $20 per million output tokens. Muse Wonder Full will not be offered through that API and is instead restricted to Muse Code subscribers, whose plans cost $50 or $200 per month, according to CNBC.

Meta said its new models include its strongest safeguards to date against cyber misuse. The company also highlighted their ability to handle extended tasks, retain user information and operate digital tools more effectively.

The launch comes as Meta faces questions about privacy and the behavior of its personal AI agent. After Muse reached the top of Apple's App Store following its release in September, some users alleged on social media that the agent accessed private messages without authorization. Meta disputed those allegations.

Wang said the company was monitoring incidents involving its AI agent and working to address user concerns. He suggested that some complaints could stem from users forgetting permissions granted earlier in a conversation.

"I think there are certainly instances users have authorized various communications or whatnot earlier in a conversation, and then may have forgotten about that," Wang said.

He added that Meta was working to improve user understanding of the agent's behavior and permissions.

"To ensure that consumers can trust the product going forward." Wang said.

Meta also said the new models represent progress following an extensive overhaul of its AI development efforts.

"Fourteen months ago, we began an overhaul of our AI efforts, and in that time we've gone from rebuilding our foundations to pushing the frontier," Meta said in a blog post.

The New York Times reported that Meta had paused a stage of model development known as pretraining in July, delaying the models' release until October. The launch represents the company's first release explicitly positioned to compete in the frontier AI category, which generally involves models requiring substantial computing resources.

Meta CEO Mark Zuckerberg had teased the release during an Instagram livestream over the weekend, mentioning a watermelon while discussing the forthcoming models. The company is seeking to expand its AI business after committing substantial resources to recruiting researchers and developing new products for its services, which include Facebook, Instagram and WhatsApp.