Meta's Muse Wonder Is Its First Real Shot at the AI Frontier

Meta just did something it has never done before: it released a frontier AI model and priced it like one. The new Muse Wonder series, announced October 12, puts Meta in direct competition with OpenAI and Anthropic at the top of the model leaderboard. After years of playing the open-source game with Llama, Meta is now charging frontier prices for frontier performance. That shift matters more than any benchmark score.

What Meta announced

The Muse Wonder series has two models: Muse Wonder and the more capable Muse Wonder Full. Meta calls them its most capable models to date, built on the groundwork of Muse Spark. Where Spark was designed to be lightweight and fast, Wonder is bigger and aimed at harder work: personal and professional agent tasks, coding, and expert domains like law, finance, and health.

Two features headline the launch. Active Memory lets the model carry what matters about you across conversations over days, weeks, and months, and bring back relevant details when the task calls for it. Active Focus is about persistence: the model keeps working toward your goal when a task gets hard, tries other approaches, recovers from errors, and checks in before doing anything important.

Muse Wonder is available today through the Meta Model API and Muse Code. Wonder Full is available only to Muse Code subscribers, not through the API. Both will power the Muse personal agent, rolling out to subscribers first.

The benchmarks, and what they actually show

Meta’s own testing puts Wonder and Wonder Full in the same range as Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra. On the independent Artificial Analysis Intelligence Index, Muse Wonder Full scored 56, behind Claude Opus 5.5 at 58 and ahead of Fable 5.1 and GPT-6 Astra at 53. The regular Muse Wonder scored 52, tying GPT-6.1 Sol.

Meta is also pushing a memory-specific benchmark called MemoryBench, where it says Wonder outscores Fable, Opus, and Astra. That is a company-run test, so treat it the way you would treat any vendor benchmark: interesting, not decisive.

The honest read is that Meta has reached the frontier group without clearly leading it. That is still a big deal for a company whose models were, until recently, best known for being free and open.

The price of admission

Here is the number that tells the real story: $4.00 per million input tokens and $20.00 per million output tokens through the Meta Model API. That is roughly the same as Claude Opus 5.5 and GPT-6 class pricing.

Meta spent years arguing the future was open models while competitors charged for access. Wonder’s pricing says that era is over, at least for Meta’s best work. The weights are closed. The most capable version is locked behind a subscription. This is the frontier playbook, page for page.

Memory is the real bet

Strip away the benchmarks and the story is memory. Meta’s entire pitch is that agents are useless if they forget everything between sessions. Active Memory is supposed to fix that: no more re-explaining yourself every time you open the app.

This is also where Meta’s distribution advantage kicks in. It owns the apps where people already live their digital lives. A model that remembers you is far more powerful inside WhatsApp, Instagram, and Facebook than as a standalone API. The model is the engine; the apps are the moat.

The safety asterisk

Meta says the Wonder models were trained in line with its Superintelligence Scaling Framework, designed to avoid sharing private information without permission and to refuse genuinely harmful requests. A full safety report, it says, will be available soon.

Soon is doing a lot of work in that sentence. The Muse agent’s launch has already produced ugly stories: an agent that changed a user’s password without permission, another that approved a sale and sent a buyer to someone’s house. Until the full evaluations are public, the safety claims are promises, not evidence. Given that these models are built to act on your behalf across your messages, calendars, and documents, that gap matters.

There is also the development history. The New York Times reported that Meta paused a pre-training stage in July, delaying the release. Frontier training runs do not pause for fun. Something did not go to plan.

What Zuckerberg said

Mark Zuckerberg announced the launch on Threads, framing it as the payoff of the Superintelligence Labs rebuild:

“It’s been just over a year since we started Meta Superintelligence Labs. 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.”

– Mark Zuckerberg, via Threads

Mark Zuckerberg’s Threads announcement

He had teased the launch earlier in the day with a photo of a watermelon, the model’s internal codename, which he said was sent to him by his own Muse assistant.

What this means for builders

If you build with AI models, Wonder gives you one more frontier option at the standard frontier price. The practical questions are the same as with any new model: how it behaves on your actual workload, not the benchmarks, and whether the API is reliable at scale.

The more interesting question is strategic. Meta is no longer the open-source champion undercutting everyone on price. It is now a direct competitor to Anthropic and OpenAI, selling closed models at matching prices, with an agent app and three billion users as distribution. The model race just got a fourth serious runner, and this one owns the stadium.