On Monday Meta unveiled two new 'frontier' AI models, Wonder and Wonder Full, which the company says perform at a level comparable to OpenAI's ChatGPT and Anthropic's Claude.

'Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier,' CEO Mark Zuckerberg wrote on X. The models will soon power Meta's Muse personal agent app for paid subscribers.

Chief AI officer Alexandr Wang called Wonder 'our largest and most capable model yet' that 'rivals the big boys.'

Meta's previous open-source Llama models lagged the industry, and its first post-restructuring model, Muse Spark, launched in April was not state of the art. A leap to claimed frontier status just months later marks a striking reversal.

According to Michael Parekh's AI: Reset to Zero, the models were trained at Meta's gigawatt-scale Prometheus data center, and the effort began from scratch only fourteen months ago with Zuckerberg's hiring spree and Wang's Meta Superintelligence Labs, built around the June 2025 $14 billion Scale AI investment that brought Wang in as chief AI officer.

Meta cited benchmarking data from the independent firm Artificial Analysis suggesting the Wonder models outperform its own Spark models and various models from Anthropic, OpenAI and Google.

According to AI: Reset to Zero, the Intelligence Index scores are 56 for Wonder Full and 52 for Wonder, up from 48 for Muse Spark 1.3 six weeks earlier, placing Wonder Full alongside Anthropic's Claude Sonnet 5.5 and ahead of OpenAI's GPT-6 Astra and Google's Gemini 4 Argon at 53.

However, Artificial Analysis's website also shows Wonder is among the most expensive models to run.

Meta has committed enormous resources to competing in AI, saying its AI-related spending could reach $145 billion this year, as it tries to diversify beyond advertising, its historic revenue base.

Meanwhile, Seeking Alpha reports Google holds a dominant 70% probability on prediction market Polymarket to have the top model by end of October, with Anthropic at 23%.

According to AI: Reset to Zero, Wonder introduces 'Active Memory', carrying important context about users across conversations, and 'Active Focus', persisting on hard goals and asking for clarification rather than stalling.

Zuckerberg's philosophy is 'the model is the product', with Muse offering no model picker or effort selector.

Meta is now one of few players with both a mainstream agent and its own frontier model, competing against OpenAI's Dots, Google's Gemini Spark, Amazon's Alexa+ and Apple's Siri AI, though startup Instinct's popularity reportedly sped up Muse's launch, per The Hindu.

The launch arrives amid industry-wide safety scrutiny. Zuckerberg said Meta delayed shipping Muse for several months to focus on safety and security, and in September he was one of six American tech executives signing a 'morally binding' commitment on safeguards, per The Hindu.

Anthropic CEO Dario Amodei told the United Nations last month that his company 'will slow down as much as necessary' to keep releases safe.

Next, Wonder rolls out to Muse paid subscribers, with Zuckerberg promising a Wonder-quality model in the smaller Spark form factor for hundreds of millions of free users.