As the tech giant pushes to diversify beyond the profits of advertising, Meta has announced two new “frontier” artificial intelligence models – codenamed Wonder and Wonder Full – that can match the performance of competing OpenAI’s ChatGPT and Anthropic’s Claude.
“Now Muse Wonder is a much bigger and more sophisticated class of models, and we are heading to a frontier,” wrote Chief Executive Mark Zuckerberg on the X platform, which is a renamed version of Twitter.
Meta’s release is the latest example of an aggressive drive under the head of Chief AI Officer Alexandr Wang, who was recruited from the Bay Area by Zuckerberg after investing $14 billion in San Francisco AI startup Scale AI in June 2025 to lead the company’s Superintelligence Labs arm.
Wonder, the new offering, follows the previous Spark offering and is built on an open-source foundation known as Llama.
Wang referred to Wonder as Meta’s “largest and most capable model yet” and “rivals the big (boys).”
Safety debates
Based on independent benchmarking from Artificial Analysis, Wonder is better than previous Meta systems, along with some of the most expensive systems on the market, from OpenAI, Anthropic, and Google.
Meta, with over 3.6 billion daily active users on its various platforms such as Facebook and Instagram, said in July it could spend as much as $145 billion on AI this year.
The release comes after Meta introduced a new research-built agentic software, called Muse Code, and released a new handheld hardware device called Muse Charm to meet the demand for independent “agentic” software that performs research, bookings and coding.
But it’s happening as there is some tension in the industry over the safety guardrails.
Zuckerberg joined executives from Google, OpenAI, Anthropic, xAI, and Nvidia in signing a White House commitment to put up safety frameworks in September.
Meta has claimed it has delayed the release of Muse because it has made improvements to its security measures, but other players have advised caution based on a number of experimental models that have either escaped from secure testing environments or performed unauthorised functions.
