Fourteen months and — by outside estimates — tens of billions of dollars after Mark Zuckerberg tore up his AI org chart and handed the keys to Alexandr Wang, Meta has a frontier model to show for it.
On Monday, Meta’s Superintelligence Labs released Muse Wonder and Muse Wonder Full, its most capable models yet, delivering state-of-the-art performance on personal and professional agentic tasks and marking the next step on its scaling ladder toward personal superintelligence.
It’s the company’s first release built to compete directly with OpenAI and Anthropic rather than carve out a cheaper, lighter lane beside them, and it arrives weeks after OpenAI shipped GPT-6 Astra on September 3 and Anthropic followed on September 22 with Claude Opus 5.5, which tells you who’s still setting the calendar in this race.
The benchmarks Meta is publishing alongside the launch undercut any claim that Muse Wonder leapfrogs the competition. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 sits at 58, Muse Wonder Full at 56, and the API-facing Muse Wonder at 52, with Claude Fable 5.1 and GPT-6 Astra both around 53. Parity with the middle of the pack, not supremacy over it. Where Meta does try to claim an edge, the numbers arrive with caveats attached: on a coding benchmark called SWE-Marathon, Wonder Full scored 57.5 versus 46.3 for Opus 5.5 and 42.5 for Astra, though Meta itself admits that row isn’t a controlled comparison and its model ran at maximum effort, meaning the flagship result came from giving Wonder Full every resource advantage while rivals were presumably held to standard settings. That’s a footnote worth reading before repeating the headline figure.
What Meta is actually betting on isn’t raw benchmark supremacy, it’s memory.
The Wonder series is built around two features the company calls Active Memory and Active Focus. Active Memory carries what a user told the model in one conversation into the next, while Active Focus keeps the model working at a hard goal, trying new approaches and checking in with the user before doing anything important.
Meta AI chief Alexandr Wang framed the design goal around agents that don’t give up the first time a task gets messy. “When you do real-world tasks, you’re going to get roadblocks and you’re going to get things that come up or various setbacks,” he said. “We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time.” For a company now building its entire AI pitch around agents that handle chores rather than chat, that’s the feature meant to carry the weight the benchmarks can’t.
The bigger break from Meta’s history shows up in how it’s charging for this. A company that built a trillion-dollar empire by giving away products and selling ads against the attention is now putting its best model behind a login and a price tag. Users of the Muse personal agent app will need to sign up for a monthly subscription of $20 or $100 to access the model, while Muse Wonder Full through Meta’s coding tool carries its own subscription tiers, reported separately at $50 and $200 a month.
The smaller Muse Wonder stays open to developers through the Meta Model API at $4 per million input tokens and $20 per million output tokens, but the flagship version doesn’t ship to everyone who wants it. CNBC put the motive plainly: the release underscores Meta’s efforts to create new AI-derived revenue sources outside of its core online ad business. Whether people who’ve never paid Meta directly for anything will start now is the open question underneath all the benchmark talk.
The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year, the physical output of the infrastructure spending Zuckerberg has spent all year defending to investors. The project traces back to Wang’s overhaul of Meta’s AI efforts in August of last year, after which Meta released its lightweight Muse Spark model and Muse agent but avoided the kind of state-of-the-art releases OpenAI and Anthropic are known for.
Muse Wonder closes that gap rather than opening a new one ahead of it, and whether that’s enough depends on what Meta is actually trying to win. A model that lands a point behind Claude Opus 5.5 on a leaderboard isn’t going to convert anyone currently loyal to ChatGPT or Claude. But an agent that genuinely remembers what you told it last week, and keeps working through a tangled insurance claim instead of giving up, might matter more to Meta’s three billion existing users than another decimal point on a benchmark chart.