Meta Launches Muse Wonder: Zuckerberg Says New AI Model Has Reached the Frontier

Mark Zuckerberg says Meta is finally back at the AI frontier. On Monday, the company launched Muse Wonder and Muse Wonder Full — the biggest and most capable AI models Meta has ever built — and declared that its expensive, fourteen-month rebuild of its AI division has pushed it to the cutting edge. “It’s been just over a year since we started Meta Superintelligence Labs,” Zuckerberg wrote on Threads. “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.” Investors liked what they heard: Meta stock jumped more than 4% on the day.

What is Muse Wonder?

Muse Wonder is the new flagship model family from Meta Superintelligence Labs, the division Zuckerberg created about a year ago after overhauling the company’s AI efforts in spring 2025 — an overhaul that followed the disappointing launch of Llama 4 and included hiring former Scale AI chief Alexandr Wang to lead the push toward what Zuckerberg calls “personal superintelligence.” Muse Spark, launched in April this year, was the lab’s first model — a lightweight one tuned for speed. Wonder goes the other direction: bigger and slower by comparison, but built for complex, high-stakes work.

The model, previously developed under the codename “Watermelon,” will power Meta’s Muse personal-assistant agent app — the AI agent introduced last month that has amassed nearly nine million global downloads. Wonder will run Muse for paying subscribers on both the $20-a-month and $100-a-month tiers. It is also available to developers today through the Meta Model API and Muse Code, the company’s AI coding tool. Wonder Full, the larger variant, is not on the API at all — it is locked to Muse Code subscribers only.

What Meta claims it can do

According to Meta, Wonder brings better long-term memory — a capability the company calls “active memory” — better focus for tackling challenging goals, and deeper expertise in law, finance and health, plus frontier-level coding performance. Meta says the model outperforms Anthropic’s Claude Fable and OpenAI’s GPT-6 Astra on several tests, and its own model card places Wonder and Wonder Full in the same range as Fable 5.1 and GPT-6 Astra. Pricing follows the frontier pattern too: through the API, Wonder costs $4.00 per million input tokens and $20.00 per million output tokens.

“We think it is competitive as a frontier model with the very best models that are out there,” Wang told The Wall Street Journal. “But in addition to that, it also is specifically the very best model for personal and professional agents and will be the best model to power an experience like Muse.” The real test, as the Journal notes, will be whether Wonder can improve retention and accelerate adoption of Muse — downloads have started to slow, and its 14-day retention lags ChatGPT, according to Sensor Tower data.

The skepticism — and the open questions

Here is the catch: Meta’s launch announcement contained no benchmark scores, no technical specifications, and no parameter counts — the standard numbers the industry uses to judge a frontier claim. “A claim of catching up with frontier rivals is a comparison, and a comparison needs two columns,” one technical analyst noted. Meta also disclosed nothing about how the models were built or trained, and while it promises larger open-weight models in the future, Wonder itself is closed and priced like its rivals — a long way from the open Llama era, when Meta’s models could be downloaded and run on anyone’s hardware.

The backdrop matters. Meta was once an AI trailblazer — Llama was the first frontier open-weight model from a big lab — but it fell behind as Anthropic, OpenAI and Google raced ahead, and Zuckerberg’s billions into Superintelligence Labs were widely seen as a gamble. Whether Wonder is truly frontier-class will be decided not by Monday’s announcement, but by independent benchmarks and real developer use in the weeks ahead. Zuckerberg, for his part, is done hedging: he told the Financial Times that Meta has caught up with rivals at the frontier — and rejects calls for a slowdown.

Follow Global Insight for clear explainers on the AI race — who is actually ahead, and who is just saying so.