Meta's Muse Wonder Is Its First Frontier AI Model, and It Marks the End of the Llama Era

Meta says it has finally reached the frontier. On Monday, October 12, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, a pair of artificial intelligence models the company calls its most capable yet, built for long-running agentic work rather than quick chat answers. The launch is Meta's first real attempt to stand beside OpenAI and Anthropic at the top of the model rankings, and it arrives just over a month after Meta introduced the Muse line itself.

The release matters for what it signals as much as for what it ships. Two years ago Meta was still the open-weights champion, giving away Llama downloads while rivals charged for API access. Muse Wonder is closed, priced at $4 per million input tokens and $20 per million output tokens, a rate that matches Anthropic's Claude Opus 5.5 rather than undercutting it. The open era of Meta AI, at least for its flagship models, is over. The subscription era has begun.

What Meta actually shipped

There are two models, with two very different access paths. Muse Wonder, model id muse-wonder-1.0, is available now through the Meta Model API and through Muse Code, Meta's coding assistant. Muse Wonder Full, the stronger variant, has no public API at all: it is locked to Muse Code subscriptions on the Power Usage and Maximum Usage tiers, aimed at developers running heavier agentic workloads.

The headline specs are frontier-grade. The model takes text, images, video, audio, and PDF as input, produces text output, and offers a context window of 1,048,576 tokens, roughly a million, enough to hold an entire codebase or a small library of documents in a single prompt. It supports tool calling, structured output, search grounding, computer use, and adjustable reasoning effort from minimal to maximum. It was trained at Meta's Prometheus data center, and it builds on the Muse Spark models Meta released in September.

Both models will eventually power Muse, Meta's consumer AI assistant. Meta says that rollout will start with subscribers before any broader release, but no date has been given.

The pitch is memory, not just benchmarks

Meta's launch materials center on two capabilities with plain names and specific jobs. Active Memory carries useful context across conversations and retrieves what is relevant, aimed at the problem of re-explaining your business every time you open a chat. Active Focus keeps working on a hard goal: trying other approaches when something fails, recovering from errors, asking clarifying questions when stuck, and checking in before important actions. These are Meta's own descriptions, not yet independently tested, and the honest read is that they target the failure mode that actually costs money in agent deployments: an agent that quits at the first obstacle or reports a task finished when it is not.

Meta says the models are strongest at agentic work in law, finance, and health, and at long-running software projects handled end to end. On safety, Meta says Wonder is its most aligned model to date, trained to avoid sharing private information without permission, to decline requests that would cause real harm, and to help defenders against cyberattacks while letting security teams probe their own code. The models are released under Meta's Superintelligence Scaling Framework, with a fuller safety report promised later.

Meta's pitch is not that Muse Wonder beats everyone on a leaderboard. It is that an agent which remembers your business and refuses to give up is the part of the stack that was missing.

The end of the Llama era

The most consequential line in the announcement may be what is missing: weights. Unlike Meta's earlier Llama releases, Muse Wonder ships closed, available only through Meta's API, Muse Code, and Meta's own apps. Meta has not said whether the weights will ever be released. The open-weight line has not vanished entirely: Muse Glimmer, a 30-billion-parameter model released in August under the Apache 2.0 license, remains open. But the company's flagship intelligence is now proprietary, which means Meta has joined the business model it once undercut.

Chief executive Mark Zuckerberg told the Financial Times that the company had reached the AI "frontier," and Bloomberg's coverage of the interview described the model as "on par with OpenAI's ChatGPT and Anthropic's Claude." The claim follows an expensive, headline-grabbing AI hiring spree that built the Superintelligence Labs team over the past year. For investors, the subtext is that the spending is supposed to convert into product. Early market coverage was upbeat: Meta shares were flagged as a notable gainer on October 12, with the stock reportedly rising more than 3 percent over the announcement and the following session.

How it scores, and what it costs

On Meta's own published numbers, Muse Wonder Full is genuinely in the conversation. In the company's 14-test table, Wonder Full beats Claude Opus 5.5 on 8 tests and loses on 6; the API version, plain Muse Wonder, beats Opus 5.5 on 5 and loses 9. Independent scoring is tighter and less flattering: Artificial Analysis put Wonder Full at 56 on its Intelligence Index, behind Claude Opus 5.5 at 58 and ahead of GPT-6 Astra at 53. The fair verdict is a frontier model, close to the leaders, not a clear winner.

Pricing tells its own story. At $4 per million input tokens and $20 per million output, Muse Wonder costs exactly what Claude Opus 5.5 charges, and far less than the $10 and $50 OpenRouter lists for GPT-6 Astra. Muse Wonder Full has no public price because it is not sold directly; it comes bundled with the higher Muse Code plans. The comparison that matters is less the per-token number than the model behind it: Meta has stopped treating frontier access as a giveaway and started treating it as a product.

What to watch

Three things will decide whether this launch is a milestone or a press release with good benchmarks. First, the Muse app rollout: when subscribers get the new model inside Meta's consumer assistant, we will learn what "personal superintelligence for everyone" actually feels like. Second, independent evaluations: Meta's numbers are Meta's numbers, and the Artificial Analysis score suggests the gap between marketing and reality is real but narrow. Third, the safety report: Meta's boldest safety claims arrive with the actual system card, which is still pending.

The broader signal is harder to miss. The open-weights champion of 2024 is now selling closed frontier models at parity prices, betting that memory and persistence matter more than raw scores. Every lab has converged on the same thesis: the next war is not about who answers best, but who finishes the job. Meta just bought its ticket.