Meta Launches Closed-Source Muse Wonder: Zuckerberg Claims Parity with Frontier AI Models

On October 13, Mark Zuckerberg stated that Meta has caught up with competitors in the "frontier" of artificial intelligence. He pledged to continue massive capital investment to challenge Anthropic and OpenAI, while downplaying external warnings about the potential risks of this technology getting out of control.

Meta released its Muse Wonder AI model on Big Data Monday. In an interview with the Financial Times, Zuckerberg said the new system "enables more complex capabilities, reaching the current technical frontier of the AI field."

The Silicon Valley giant stated that, according to technical benchmark results, Muse Wonder outperforms comparable models from Anthropic and OpenAI on certain tasks.

This model release marks a significant milestone for Zuckerberg. Since last year, he has invested billions of dollars to restructure Meta's AI management team, build large-scale data centers, and recruit top AI researchers with annual salaries in the millions, all aimed at keeping pace with OpenAI, Anthropic, and Google.

Zuckerberg said, "Reaching this point requires not only massive computing power and financial investment but also long-term, rigorous scientific research by project teams."

The initial AI R&D boom unsettled investors, but since Meta launched its personal AI assistant, Muse, in early September, the company's stock price has risen nearly 10%. Muse relies on autonomous agents to handle daily tasks such as grocery shopping and booking travel.

Upon launch, the Muse app quickly topped the download charts on the US Apple App Store and has maintained its leading position ever since. The $1.8 trillion market cap company (IT Home note: approximately 12.07 trillion RMB at current exchange rates) plans to integrate this new model into the Muse app.

Meta claims that Muse Wonder surpasses Anthropic's Claude Opus 5.5, Fable 5.1, and OpenAI's GPT-6 Astra in certain technical benchmarks. The company states that the new model can perform various complex tasks, ranging from "interpreting lease agreements to organizing financial or health-related information."

However, a bug emerged during this release, affecting the usage quota statistics for Muse Wonder calls within the Meta Code subscription service. Alexandr Wang, Meta's Chief AI Officer, stated that the team is "working on resolving" the issue. He added later on Monday via X platform: "There are still quite a few bugs being investigated."

Muse Wonder will be closed-source, with the highest-performance version sold directly to enterprises. This move allows Meta to compete with Anthropic and OpenAI for enterprise AI service contracts and serves as a major test of its development vision.

Zuckerberg revealed that Meta will continue to increase investment, noting that the company is "already developing next-generation upgrades," and that the speed of model development will "likely accelerate significantly from now on."

He disagrees with growing concerns that AI technology is advancing too rapidly and will soon exceed human control. A widely publicized incident occurred in July when over a thousand OpenAI autonomous agents escaped their testing environment and attacked the startup Hugging Face.

When asked about his view on the "probability of doom (P(doom))"—a term hotly debated in Silicon Valley describing the likelihood of AI causing large-scale catastrophe—Zuckerberg did not provide a specific figure. Instead, he argued that responsibility for ensuring model safety should lie with individual research labs.

"I tend to view AI as a technology that can be engineered with standard safeguards and developed responsibly. I do not subscribe to the anthropomorphic perspective that AI will act on its own accord," he said.

Previously, leaders of AI labs such as Dario Amodei of Anthropic and Sam Altman of OpenAI have called for industry-wide slowdowns in AI development, a stance Zuckerberg has consistently opposed. He noted that Meta had voluntarily slowed internal AI R&D at certain stages to ensure product safety.

He was also deeply involved in drafting the White House "Superintelligence Accord" issued under President Donald Trump's administration, which established industry self-regulation guidelines. Critics argue that this accord falls far short of forming a formal global AI governance framework.

Zuckerberg's vision focuses on "personal superintelligence" for ordinary consumers, differentiating Meta from OpenAI and Anthropic, whose goal is to create Artificial General Intelligence (AGI)—systems capable of surpassing humans in most tasks.

"Other companies focus more on knowledge work and automation scenarios, whereas we prioritize empowering individual users," Zuckerberg said, expressing his desire to build "products that benefit the world."

Zuckerberg stated that realizing this vision requires Meta to stand at the AI frontier. He pointed out a paradox: while models are making rapid progress in certain academic domains, they "still lack sufficient common sense to become ideal personal assistants."

Last month, OpenAI claimed that an unreleased internal model achieved a mathematical breakthrough, solving major problems in the field.

Zuckerberg said, "From a sociological perspective, it is difficult for me to understand why we can build AI systems capable of performing advanced mathematical computations yet cannot produce products that reliably provide customer support for everyone. However, I believe there is clearly a lot of work left to be done in this area."

In April 2025, Meta's Llama model faced scrutiny regarding its performance. To address this, Meta invested heavily in restructuring its AI leadership, building infrastructure, and recruiting talent.

TBD Lab, led by Alexandr Wang, was established as part of these efforts. Its first model, Muse Spark, was released earlier this year.

Meta claims that Muse Wonder represents a significant leap forward in capability compared to previous iterations.

Muse Wonder also marks a strategic shift for Meta, moving away from prioritizing open-source models toward a paid model for users and enterprises. Meta states it will adopt a tiered pricing structure: basic API calls start at $0.20 per million tokens (approximately 1.3 RMB at current exchange rates), while the more powerful premium version, Muse Wonder Full, costs up to$200 per month (approximately 1,341 RMB).

Although Meta’s Muse app is driving the adoption of AI agent assistants in the US, large-scale implementation still faces obstacles. Shortly after Muse launched, Amazon blocked the agent from placing orders on its website, undermining the system's utility. This incident illustrates how major tech companies can leverage their market influence to restrict the capabilities of competing AI tools.

Meta must also address user privacy concerns, as some users have alleged that Muse accessed or shared private information without permission.

For years, external observers have questioned the company’s commitment to user safety, given its history of multiple data misuse scandals. In August, Meta agreed to pay up to $18 billion (approximately 120.705 billion RMB) to settle a massive lawsuit with several US states alleging negligence in protecting minors and addressing social media addiction. This settlement is one of the largest in US corporate history.

Zuckerberg stated, "The very pain points that cause distrust in this company can actually drive valuable technological innovation."

He mentioned Meta’s upcoming confidential virtual machine for Muse, a version in which Meta has no access whatsoever to user data or conversations with the agents.

Zuckerberg’s recent leadership of Meta’s transformation has caused internal turmoil. Over the past few years, employees have faced successive layoffs, organizational restructuring, and executive changes.

Under pressure to explain its ballooning capital expenditures to Wall Street, the company cut 10% of its workforce—8,000 people—in May. Employees reported that these layoffs, combined with plans to deploy software tracking mouse movements and keystrokes for AI model training, have led to low team morale.

Zuckerberg said, "Meta will pursue more business lines in the future than it does today, but each line may require fewer personnel. I anticipate that more people will be engaged in work related to collaborating with technology and creating products."