Meta's first frontier model: Muse Wonder launches — with the top version locked behind a subscription
Meta's Superintelligence Lab on Monday, October 12, 2026, released two new models, Muse Wonder and Muse Wonder Full — the company's first attempt at a frontier model. But the strongest variant, which also scores best on independent benchmarks, is not available through the API: it is reserved for subscribers of Meta's own coding tool, while privacy complaints about the Muse app remain unresolved.
What happened
On Monday, Meta launched Muse Wonder and Muse Wonder Full, which according to TechCrunch is the first time the company attempts a frontier model — that is, a model intended to compete at the top tier alongside OpenAI, Anthropic and Google. The launch was followed the same day with pricing for both API access and consumer subscriptions, CNBC reports.
Meta AI chief Alexandr Wang called the models frontier models in an interview with Bloomberg: "These are true frontier models," he said, according to Bloomberg via MSN. "We really think they are competitive with the best models out there, and particularly state-of-the-art on personal and professional agentic work." Wang came to Meta through a deal worth $14 billion tied to Scale AI.
How the models score — and who is measuring what
The picture of how good the models are depends on who is asking the question.
According to Meta's own testing, presented in the model card, Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra, TechCrunch writes. The FT, carried via Yahoo Finance, reports that Meta says Muse Wonder "outperforms" Claude Opus 5.5, Fable 5.1 and GPT-6 Astra "on some technical benchmarks." These are the company's own measurements, in other words.
The independent benchmarking firm Artificial Analysis places the models differently. On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scored 56 points at maximum reasoning effort, according to Cryptobriefing. That makes it the second-highest-ranked model on the index — behind Anthropic's Claude Opus 5.5. Standard Muse Wonder scored 52, up from predecessor Muse Spark 1.3's 48.
The difference between the two pictures is significant: Meta's "outperforms on some benchmarks" and an overall second place on an independent index are not the same claim. Readers comparing providers should therefore distinguish between the company's own figures and third-party measurements.
Access and pricing: the strongest model behind the subscription wall
The pricing structure is perhaps the most telling part of the launch.
Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens — roughly on par with GPT-6 Sol or Claude Opus 5.5, TechCrunch writes. But Wonder Full, the variant that scores highest on the independent index, is not offered through the API at all. It is available only to developers using Meta's coding-focused agent Muse Code, under subscription plans of $50 or $200 per month, CNBC reports.
For regular users, a monthly subscription of $20 or $100 is required to access Muse Wonder through the Muse app, Meta's personal agent. CNBC notes that the launch underscores Meta's effort to create new revenue streams outside its core business — advertising. The fact that the model's safety features, as described below, currently come only with the paid model reinforces the impression that commercialization is central.
For developers, this means a practical limitation: anyone who wants to build on the strongest model via the API cannot do so. They must either settle for standard Wonder or subscribe to Meta's own agent product.
"Active memory" as the differentiator
Meta highlights the models' memory as the most important competitive advantage. According to the company's own testing, Muse Wonder now scores higher on MemoryBench than Fable, Opus or Astra — a claim that comes from Meta's measurements, not independent verification.
Research scientist Kavya Kopparapu explained the work in a video released alongside the launch: "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months and beyond," she said, according to TechCrunch.
Agentic reliability is the second pillar. Wang told CNBC that Meta has trained the model not to get stuck: "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," he said. The idea is that a personal assistant should be able to go longer on its own before needing the user's input.
Safety promises and privacy doubt
Meta wrote in a blog post that the new model "offers our strongest safeguards to date" against "cyber misuse." CNBC points out, however, that these protections currently come only with the paid model.
At the same time, privacy questions shadow the Muse app. Several users have complained on social media that the agent allegedly accessed personal messages without authorization, CNBC writes — but the company has rejected the complaints. Newsbytes reported the same day that the Muse app has climbed Apple's App Store charts, but that there have been user concerns about message security, and that Meta is working on better user education.
In response to the concerns, Meta plans a feature called "Muse Confidential VM," according to Bloomberg's coverage via MSN, which is intended to shield private user interactions from access by the company itself. It is worth emphasizing: the complainants are user reports on social media that Meta disputes, and there is no independent investigation in the available source material. The safety claims in the blog post are similarly the company's own assertions, not independently verified.
Market reaction and strategic framing
Meta's stock rose as much as 4.7 percent after the announcement, according to Bloomberg's coverage via MSN, which describes growing investor enthusiasm for the company's aggressive capital expenditure. The picture is nonetheless not uniform: the same source reported that retail sentiment is measured as "bearish," showing that the two indicators capture different things.
In an interview with the Financial Times, Mark Zuckerberg rejected calls for an industry-wide slowdown of AI development: "I'm much closer to the belief that AI is a technology that can be engineered with normal guardrails and built responsibly than the anthropomorphic vision of AI, which is that it is just going to go do its own thing," he said, according to the FT via Yahoo Finance. He further signaled continued investment and faster model development ahead.
The open questions
Several matters remain unresolved after launch day.
First, the gap between Meta's benchmark claims and the independent index: if Wonder Full "outperforms" Opus 5.5 on technical benchmarks, why does it score lower on the overall index? The answer may lie in the fact that these are different measurements, but that remains speculation for now — the sources do not explain the gap.
Second, the privacy complaints, which have been neither confirmed nor refuted independently. How "Muse Confidential VM" will work in practice, and when it will arrive, is also not specified in the available material.
Third, the access question: it is unclear why Meta is keeping Wonder Full off the API, and whether this is a permanent strategy or a temporary move. No architecture or weight details have been made public either.
What is certain is that with Muse Wonder, Meta is entering the frontier class for the first time — with measurable progress on an independent index, a deliberate subscription strategy around its strongest model, and a number of claims still awaiting independent verification.