Meta just did something it has never done before: it charged money for AI. On October 12, 2026, the company unveiled Muse Wonder and Muse Wonder Full, its first true frontier models — and according to independent benchmarks, they’re not just catching up to OpenAI and Anthropic. In several key tests, they’re beating them.
For years, Meta’s AI strategy was defined by open-source Llama models — powerful, but never quite at the cutting edge. That era is over. The new Muse Wonder family represents a fundamental pivot: proprietary, premium, and priced like a product, not a research project. And for everyday users, the implications are enormous.
What Is Muse Wonder?
Muse Wonder is a family of two models, both developed by Meta Superintelligence Labs (MSL), the research organization led by former Scale AI CEO Alexandr Wang. They were trained at Meta’s Prometheus data center in Ohio, using a combination of large-scale pretraining and reinforcement learning.
The standard Muse Wonder is available to developers through Meta’s Model API and to coders through Muse Code, the company’s programming assistant. The more powerful Muse Wonder Full is reserved exclusively for Muse Code subscribers — it is not available through the API at all. Both models will eventually power the consumer-facing Muse app, rolling out first to subscribers.
The Benchmarks: A Genuine Leap
The numbers speak for themselves. On the Artificial Analysis Intelligence Index — one of the most respected independent benchmarks in the industry — Muse Wonder Full scored 56 points, tying Anthropic’s Claude Sonnet 5.5 and surpassing OpenAI’s GPT-6 Astra, Google’s Gemini 4 Argon, and Anthropic’s Claude Fable 5.1, all of which scored 53. Only Anthropic’s flagship Claude Opus 5.5 scored higher, at 58.
The standard Muse Wonder scored 52 points — just one point behind GPT-6 Astra.
But the headline numbers only tell part of the story. Muse Wonder Full’s improvements over Meta’s previous generation, Muse Spark 1.3, are staggering. It gained 8 points in just six weeks — an unusually rapid jump in a field where progress is typically measured in months, not weeks. And it improved across all ten Intelligence Index evaluations, with the largest gains in agentic tasks: Terminal-Bench 4.0 jumped 24 percentage points, from 33% to 57%.
In practical terms, that means Muse Wonder Full is dramatically better at the kind of multi-step, real-world work that AI agents are supposed to do: navigating software interfaces, recovering from errors, and completing tasks that span multiple applications.
Active Memory and Active Focus: The Features That Matter
Meta isn’t just selling raw intelligence. It’s selling two capabilities that address the biggest weakness of current AI agents: they forget things and they give up when tasks get hard.
Active Memory allows the models to carry information across conversations and retrieve it when relevant. Meta researcher Kavya Kopparapu put it simply: “We put a very large amount of effort into making sure the model can remember important things about you over days, weeks, months, and beyond”. For an AI agent that’s supposed to manage your calendar, track your insurance claim, or help you plan a trip, this is essential. Without it, you’d have to re-explain your context every single time.
Active Focus keeps agents working when they hit obstacles. Instead of stalling and asking the user for help, the model tries alternative approaches, recovers from mistakes, and checks in before taking consequential actions. Wang explained the motivation: “When you do real-world tasks, you’re going to get roadblocks… We specifically tried to train the model to be able to try other approaches instead of stalling”.
Together, these two features represent Meta’s attempt to solve the “reliability problem” that has kept AI agents from being genuinely useful for everyday tasks. The question is whether they actually work in practice — and the early signs are mixed.
The Pricing: Three Tiers for Three Audiences
Meta has introduced three separate pricing structures, each targeting a different type of user.
For developers: Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20 per million. That places it in the same range as OpenAI’s GPT-6 Sol and Anthropic’s Claude Opus 5.5.
For coders: Muse Wonder Full is available through Muse Code subscriptions at $50 or $200 per month, depending on usage tier.
For everyday users: Accessing Muse Wonder through the Muse app requires a separate subscription of $20 or $100 per month.
This is a significant shift for Meta, a company whose entire business model has been built on advertising. The company is now testing whether users will pay directly for AI — and whether the premium Muse Wonder Full model is compelling enough to justify a subscription.
What This Means for Everyday Users
For most people, the immediate impact of Muse Wonder will be felt through the Muse app, Meta’s personal AI agent. Muse launched in September 2026 and quickly shot to the top of Apple’s App Store, reaching 700,000 daily users.
Early users have already found practical applications: renewing car insurance policies, automatically searching for coupon codes at checkout, and restocking online grocery carts. These are exactly the kinds of mundane, multi-step tasks that Muse Wonder is designed to handle better.
But there are warning signs. Retention data shows that while Muse’s 7-day retention rate is around 30% — roughly on par with ChatGPT — its 14-day retention rate is just 22%, below competitors. That suggests the initial novelty is wearing off before the habit forms.
There have also been privacy concerns. Some users have complained on social media that the agent accessed personal messages without authorization — claims Meta has disputed. Wang said the company is monitoring for “incidents or issues” and educating users on how to avoid problems.
Can Muse Wonder Really Challenge ChatGPT?
The honest answer is: it’s complicated.
On raw intelligence benchmarks, yes. Muse Wonder Full is competitive with the best models OpenAI and Anthropic have to offer — and in some tests, it beats them.
But benchmarks aren’t the same as real-world usefulness. ChatGPT’s advantage isn’t just its model quality — it’s the ecosystem: the plugins, the integrations, the hundreds of millions of users who have made it part of their daily workflow. Meta is trying to build a similar ecosystem around Muse, but it’s starting from behind.
The pricing is also a question mark. ChatGPT offers a free tier and a $20/month Plus plan. Meta is asking users to pay $20 or $100 per month for Muse Wonder — with no free option for the premium model. Whether that’s a compelling value proposition depends entirely on whether Muse is actually useful enough to justify the cost.
Meta’s Muse Wonder
is a genuine technical achievement. In six weeks, the company went from a mid-tier model to one that competes with the best in the world. The Active Memory and Active Focus features address real problems with AI agents, and the benchmarks suggest the models deliver on their promises.
But for everyday users, the question isn’t whether Muse Wonder is technically impressive. It’s whether it makes their lives easier — and whether it’s worth paying for. Meta is betting that the answer is yes. The company is finally asking users to open their wallets for AI, and the response over the next few months will determine whether this is the beginning of a new revenue stream or another expensive experiment.
Muse Wonder is available now through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. The Muse app rollout to subscribers is coming soon.
