Meta’s Muse Wonder Overtakes GPT-6 and Gemini 4
Trending Topics Jakob Steinschaden ● Covered by 4 sources
Meta says its new Muse Wonder models beat OpenAI and Google in tests. But the big one isn’t open yet, so the win comes with a paywall.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mark Zuckerberg just spent a mountain of cash on people and data centers, and now Meta has a number to defend. The company has unveiled two new A.I. models, Muse Wonder and Muse Wonder Full, and says both can match or beat the latest flagship systems from OpenAI and Anthropic on its own tests.
The models were built under the internal name Watermelon and trained in Prometheus, Meta’s gigawatt-scale data center. Meta is leaning hard on two agent-focused features: Active Memory, which keeps track of useful details across conversations, and Active Focus, which is meant to help an agent stay on task, try different routes and recover when things get messy. The company also says the models are strong at coding and at specialized work in law, finance and medicine.
The claims are not just Meta talking into the mirror. Artificial Analysis, an independent benchmarking firm, puts the strongest version near the top of its Intelligence Index, which rolls ten tests together from coding to agentic office work. Muse Wonder Full scores 55.8 points, behind Anthropic’s Claude Opus 5.5 at 57.6, but ahead of GPT-6 Astra at 52.7 and Gemini 4 Argon at 52.6. Meta’s standard Muse Wonder lands at 52.4, just behind those two. On the same index, Meta’s previous best model, Muse Spark 1.3, scored 48.
That makes this more than a routine model drop. After the rough launch of Llama 4 in spring 2025, Zuckerberg reshuffled Meta’s A.I. operation, hired Scale AI founder Alexandr Wang as chief A.I. officer, and threw huge pay packages at researchers. Wang told Axios these are “true frontier models.” Now the pressure shifts from promises to proof.
And there’s the awkward part: the best version is closed. Muse Wonder and Muse Wonder Full run only through Meta’s own API and subscriptions, with no weights released. Meta says a model card is coming and a detailed technical report will follow, but it has not said whether an open version will arrive. For a company that used to talk like open A.I. was a moral cause, this is a very familiar modern ending: open for the small stuff, closed for the crown jewels.
My take — AI-written commentary, not fact-checked reporting
Meta has joined the club it used to sneer at: frontier model makers who keep the good stuff locked up and call it strategy. The open-source sermon is still there, but the weights are not. That’s not hypocrisy so much as a business model wearing a hoodie.
Read more about this at: Trending Topics