Meta announces the launch of all-new AI models Muse Wonder and the more powerful Wonder Full, claiming their capabilities have reached industry frontier levels. According to test data released by Meta, these two models have approached the level of Fable 5.1, Opus 5.5, and GPT-6 Astra in multiple capability tests, and slightly outperformed them in certain agent and knowledge work benchmarks. This move comes as Anthropic and OpenAI have successively called for the industry to slow frontier AI large model development.
Meta's new AI models' capabilities reach industry frontier level
As background, after Meta's Llama 4 large model launched in April last year suffered a setback, Zuckerberg conducted a large-scale reorganization of the development team, bringing in Tao Wang, then 28, as Chief AI Officer. In April this year, the reorganized Meta Superintelligence Lab released its first model, Muse Spark, focusing on low latency and lightweight operation. Over the following months, Meta iterated on the Spark model multiple times, culminating in this week's release of the Wonder model, which has reached industry frontier level.
Notably, the "personal superintelligence" vision pursued by Zuckerberg differs markedly from the AGI pursued by the other two giants. The two capabilities emphasized by Meta Wonder this time include "proactive memory" and "sustained focus." The former enables the model to retain important context across conversations and proactively retrieve relevant information when handling new tasks; the latter allows it to persist in the face of complex goals, try alternative approaches when encountering obstacles, and adjust plans based on new information.
Zuckerberg emphasizes AI technology safety and industry self-discipline
This also corresponds to a practical need for personal AI assistants: users need not reintroduce their situation each time, and the assistant can also sustain work requiring multiple rounds of operation. Meta's demonstrated cases include formulating running plans across apps based on personal preferences, designing and ordering customized gifts using real photos from personal social media pages, and arranging wedding seating according to complex matching requirements.
In Meta's internal MemoryBench evaluation, Muse Wonder scored 85.6, higher than GPT-6 Astra's 81.4 and Opus 5.5's 81.5. The test required models to retrieve key information from multi-turn conversations spanning 64,000 to 100,000 tokens, while coping with interference from irrelevant content and conflicting or subsequently updated information. In programming, Meta emphasized the model's ability to track project requirements, modify codebases, and continuously conduct testing and revision.
Its published SWE-Marathon results show Wonder Full achieved a 57.5% pass rate in the evaluation examining ultra-long software engineering tasks, while Wonder scored 53.1%, both higher than the three competing models in the comparison table. As model capabilities improve, corresponding prices have risen accordingly. Muse Wonder is charged via API by usage, with standard input, cached input, and output priced at US$4 (approx. HK$31), US$0.20 (approx. HK$2), and US$20 (approx. HK$156) per million tokens respectively; the more powerful Wonder Full is offered through the programming tool Muse Code, included in monthly subscription plans of US$50 (approx. HK$390) and US$200 (approx. HK$1,560).
At Monday's promotional event, Zuckerberg hinted that Meta would further increase investment, stating the company "is already working toward the next level." He also said the pace of model development going forward "may feel noticeably faster." However, regarding growing external concerns that the technology's development speed will cause it to exceed human control, Zuckerberg disagreed. He argued that responsibility for ensuring model safety should be borne by each lab individually, and stated "I tend to believe AI is a technology that can be designed through conventional safety safeguards, rather than that kind of anthropomorphic AI imagination."
Zuckerberg previously opposed calls from AI lab leaders including Anthropic head Dario Amodei and OpenAI head Sam Altman for the industry to "slow down" AI development. He played a key role in formulating the White House "Superintelligence Accord" under U.S. President Trump, which established industry self-discipline principles.
Zuckerberg revealed some behind-the-scenes stories of Meta's AI team reorganization on the tech podcast Cheeky Pint that day. Zuckerberg said that at a year-end executive team meeting, he asked Wang whether, if the company were to invest 100 billion or 200 billion dollars next, he was certain it would succeed. Wang recalled that at the time he could only give a "vibe report" on research progress, and could not determine whether it would succeed.
Additionally, Zuckerberg disclosed that operating a one-gigawatt data center costs approximately 15 billion dollars annually, while upfront construction capital expenditure requires approximately 50-60 billion dollars. Regarding how to obtain returns on massive AI investment, Zuckerberg said Meta plans to expand Muse's user base through free or low-cost services, attract more partners to integrate, and create network effects. On this basis, the company can achieve monetization through channels such as premium feature subscriptions, commercial service revenue sharing, model APIs, and enterprise services.
He said: "If you want to start using Muse Wonder in Muse right away, all you have to do is pay for a subscription. I think this will be a big business."