Mark Zuckerberg is betting Meta can overtake OpenAI and Anthropic by building AI that people trust enough to use for real work. The company launched Muse Wonder and Muse Wonder Full on October 12, claiming a return to the frontier after more than a year spent rebuilding its artificial intelligence operation.
The launch gives substance to his competitive ambitions, but does not establish that Meta has beaten either rival across the board. It also brings renewed relevance to his September 24 rejection of a coordinated industry slowdown: Zuckerberg wants individual labs to address their own safety problems rather than wait for a collective agreement.
Meta takes its fight into coding and personal agents
The October 12 release targets coding and tasks involving legal documents, financial information and health. Muse Wonder is available through the Meta Model API and Muse Code. The more powerful Full version is available to Muse Code subscribers, while integration into the consumer Muse assistant is coming later, beginning with subscribers.
Alexandr Wang, who leads Meta Superintelligence Labs, told Axios:
These are true frontier models, competitive with the best models in the world, and are particularly state of the art on personal and professional agentic work.
That distinction matters. Developers can test the standard model now, but the strongest version is not being offered through the same API channel. A coding subscription and a consumer assistant are different products, with different permissions and practical uses.
Two capabilities underpin the pitch. Active Memory retrieves relevant details from earlier conversations. Active Focus helps an agent persist through difficult tasks, try alternatives and recover from errors, while asking for clarification and checking before important actions.
The commercial prize is software that can finish useful work rather than simply produce convincing answers. Handling a lease or insurance claim also means accessing sensitive information, making permission controls part of the product's usefulness.
Zuckerberg says labs should police their own pace
In his September 24 interview with NBC News's Joanna Stern, Zuckerberg rejected industrywide coordination while allowing that developers should slow their own projects when necessary:
I don't think that we need some kind of industrywide coordination. I think that each lab needs to take the time, and when it sees that there are issues, you just take the time that you need internally to basically make sure that you're proceeding safely.
His reasoning rests on customers abandoning unreliable agents and companies facing liability when their products cause harm. In a September 15 post, he pointed to a delay in releasing Muse:
Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us.
He also backed independent evaluators and directing most computing resources toward serving users rather than pursuing recursive self-improvement. His position therefore accepts safety work, but rejects making cooperation among competing executives a prerequisite for progress.
The safety argument is getting harder to settle internally
As previously reported by Apex, Bill Gates has called for stronger public safeguards, arguing that cybercrime and biological misuse require oversight beyond company promises. Those risks can affect people who never chose to use an AI product, limiting what customer dissatisfaction alone can resolve.
Apex had earlier reported that AI insiders were preparing for a catastrophic cyber incident and its political aftermath. OpenAI said its exercises explored possible scenarios rather than treating disaster as inevitable. The disagreement is about how to manage uncertainty, not a confirmed countdown to catastrophe.
For Wonder, the published research describes training against unauthorized sharing of private information and harmful misuse, while preserving defensive cybersecurity capabilities. It also says additional safety evaluations will be released in a full report later.
That leaves Zuckerberg's bet facing two tests: whether the new models perform reliably outside the company's selected evaluations, and whether their safeguards hold when agents gain access to real accounts, documents and decisions. Winning users from OpenAI and Anthropic will require both.