$topblogs
Artificial Intelligence

Meta finally has a frontier model. Now it wants you to pay for it.

Hanzla Baig
Hanzla Baig

October 13, 2026 · 11 min read

Meta finally has a frontier model. Now it wants you to pay for it.

Meta released Muse Wonder and Muse Wonder Full on Monday, the first models from its Superintelligence Labs that the company is willing to call frontier-class. They are bigger than anything in the Muse Spark line, they are aimed squarely at AI agents rather than chat, and most of the people who want to use them will have to pay.

That last part is the quiet headline. Meta built its reputation in part on Llama's open-weight releases and made its consumer AI assistant available without a paid subscription. Wonder follows a different model: API access is metered by token, and Wonder Full is restricted to Muse Code subscribers. That matters because Meta is now asking developers and users to pay for capabilities it presents as competitive with the leading proprietary models, and its own scorecard says Wonder is competing with the best from OpenAI and Anthropic. Whether the independent numbers back that up is the more interesting question, and the answer so far is "close, but not first."

What Meta actually released

There are two models. Muse Wonder is the broadly available one. Muse Wonder Full is the larger variant, and it is locked inside Meta's Muse Code coding agent. Meta describes both as its most capable models to date, built for what it calls personal and professional agentic tasks: the kind of work where a model has to remember who you are, use tools, and keep going when something breaks.

According to Meta's model card, both models accept text, image, video, and audio as input and produce text as output, with a context length of 1 million tokens at inference. Meta has not disclosed parameter counts. The company says they were trained at Prometheus, its gigawatt-scale data center in Ohio, and TechCrunch notes the facility came online earlier this year. Internally the project was codenamed Watermelon, according to Bloomberg and CNBC.

Meta began a ground-up rebuild of its AI organization 14 months ago, which included a roughly $14 billion investment in Scale AI that brought its founder, Alexandr Wang, in to run the new unit, per Bloomberg. Until now the output of that overhaul was Muse Spark, a lighter, lower-latency model. Wonder is the first time Meta has gone after the top of the leaderboard. "These are true frontier models," Wang said, as quoted by Quartz.

Availability and pricing: the confusing part

Meta has split access into several tracks, and the coverage is not always consistent about them, so here is the version that holds up across Meta's blog post, TechCrunch, and CNBC.

  • Meta Model API: Muse Wonder only. It costs $4.00 per million input tokens and $20.00 per million output tokens, which TechCrunch says puts it roughly level with Claude Opus 5.5. Artificial Analysis independently lists the same prices and a 95 percent discount on cached input, which works out to about $0.20 per million tokens. For comparison, price-tracking sites citing OpenAI's pricing page list GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens, roughly half Wonder's rates. Those are token prices, not a guarantee that every task costs half as much.

  • Muse Code: Muse Wonder is available here, and Muse Wonder Full is limited to paying subscribers. CNBC reports the plans cost $50 or $200 a month.

  • The Muse app: Meta says Wonder will "soon" power experiences in Muse and will reach subscribers first. CNBC reports that app users will need a $20 or $100 monthly plan to get it. Meta has not said when free users might get access, or whether they will.

The detail developers will care about: Wonder Full is not served through the API. Per TechCrunch, it exists only inside Muse Code. If you are an enterprise team shopping for a non-OpenAI, non-Anthropic frontier endpoint, you get the smaller of the two models.

Meta has not said when free Muse users might get Wonder.

What Active Memory and Active Focus are supposed to do

Meta organizes the pitch around two named features.

Active Memory is the model carrying important context about you from one conversation to the next and pulling the right details back up when a task calls for them. Research scientist Kavya Kopparapu said in Meta's launch video that the team put a great deal of effort into making the model remember what matters over days, weeks, and months.

Active Focus is persistence. Meta says the model is trained to try alternative approaches when it hits a blocker, to adapt as new information comes in, and to ask for clarification or check in before important actions. Wang gave CNBC the plain-English version: real-world tasks throw up roadblocks, and the model is meant to find another route instead of stalling and pinging you every time.

The demos on Meta's page are the usual mix of everyday and slightly improbable: a running plan built from personal preferences, a wedding seating chart, custom merchandise ordered from photos on an Instagram profile, a 3D kitchen game. They show the intended use, but they are demos, not evidence.

The benchmarks, and how to read them

Meta's launch materials compare Wonder and Wonder Full with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5, all at maximum reasoning effort. TechCrunch's summary of that testing is that the Wonder models land at comparable scores to Fable 5.1 and Astra.

Independent numbers are less flattering. On the Artificial Analysis Intelligence Index (v4.3), Muse Wonder scores 52 and Muse Wonder Full scores 56, up from 48 for Muse Spark 1.3, which shipped only about six weeks earlier. Claude Opus 5.5 still leads that index at 58, as reported by AI Weekly and Crypto Briefing. Artificial Analysis says Wonder Full placed in the top three on six of the ten evaluations in the index, and that it did best on agentic and coding tests.

Coding tells a similar story. In Artificial Analysis's Coding Agents Index, Wonder Full scored 68 when running inside Muse Code, matching Claude Code with Sonnet 5.5 (68) and scoring above the tested Opus 5.5 configuration (66). Standard Wonder scored 63. These results describe specific agent setups, not a universal ranking across every task or deployment.

Several details complicate the picture.

Meta chose the conditions. Every figure is at max reasoning effort, and Meta's own methodology notes that for some benchmarks it used the highest of its internal run, a self-reported score, or the official leaderboard number. Its competitor runs are "best-effort," in its own words, and the model card admits prompts and tools may not be tuned for rival models. The SWE-Marathon comparison also mixes different agent harnesses, so Meta itself says it is not a like-for-like test.

Some of the best numbers are Meta's own tests. MemoryBench, the headline result for Active Memory, is an internal benchmark. Meta says Wonder beats Fable, Opus, and Astra on it. Decrypt's reading of the chart puts both models around 85.5 percent. That is a reasonable result, but it is Meta grading its own homework on the exact feature it is selling.

Cost and speed are not free. Artificial Analysis puts Wonder at $4.52 per task on its Intelligence Index, about 2.8 times what Muse Spark 1.3 costs, driven by higher token prices and heavier input use. Its API tests recorded output of roughly 88 tokens per second, slightly above the tier average, with a time to first token of about 19 seconds. Those are measurements from the evaluator's maximum-reasoning configuration, not a promise about how fast every everyday request will feel, but they show why a benchmark score has to be read alongside cost and latency. Nobody has published results for Wonder at a normal effort level on a normal budget yet.

Bloomberg put the broader caveat well: benchmark results can be engineered to look favorable and often diverge from how a model behaves in practice. The honest summary is that Wonder looks like a real step up for Meta and sits in the same conversation as the leaders. It does not, on the independent aggregate, beat them.

The safety card is more candid than the blog post

Meta's blog says Wonder has its strongest safeguards to date against chemical, biological, radiological, nuclear, and cyber misuse, and calls it the company's most aligned model. The model card adds some detail that does not make it into the headline.

Under Meta's own framework, the company says it cannot rule out Wonder Full reaching the "high" risk threshold in chemical and biological capability without mitigations. It applied safety post-training and system-level classifiers, and says that with those in place residual risk is "moderate or lower" across all domains. A full safety report is promised "soon," not published today.

One footnote is worth reading closely. On the ExploitGym cyber evaluation, Wonder Full captured 92 flags, but a verification pass found only 21 used the intended vulnerability. The other 71 came through unintended ones, so the raw flag count is not a count of successful exploits. Meta reports the 21 as the figure comparable to the published ranking, which is the correct call, and also a good illustration of how much a single headline number can hide.

The card lists limitations too: visual and spatial reasoning, especially 3D, remains weak, performance may drop outside the strongly supported languages, and the model is not intended for people under 18. For agentic use it recommends human confirmation before irreversible actions.

The part that is not about benchmarks

An assistant that remembers everything about you is only as good as the company holding that memory, and Meta's track record with the Muse app is already being questioned.

Fast Company reported last week that some early adopters are deleting or restricting Muse after privacy scares. Meta says Muse runs in isolated virtual machines, keeps credentials separate from the model, asks before sending email or making purchases, and lets people disconnect apps or permanently delete their data. Wired reported, based on instructions an independent researcher extracted from the app, that Muse is told to maintain a page for every person in a user's life, including friends, family, and colleagues who may not use the product (Gizmodo and Cybernews summarize the findings). Those instructions show how the product is designed, not how many profiles actually exist. Meta's response is that a useful agent needs context about you and the people you deal with, and a spokesperson told Wired the agent uses publicly available information that users chose to share.

Wonder is explicitly designed to remember more, for longer, and to act with more autonomy. That makes the permissions model and the data-handling terms matter as much as the benchmark table. Until the full safety report lands, those claims rest on Meta's word.

Who this is for

For developers, Wonder is a credible new option at a familiar price, with the caveat that the stronger Full model is not available over the API. For consumers, the pitch is an agent that gets less stuck, but it is a paid feature, and it arrives while the app's privacy record is still being argued over. For Meta, it is the first real evidence that the 14-month rebuild produced something competitive, and it comes with an unusually direct attempt to earn money from AI outside the ad business.

The open questions are specific. Comparable independent results at lower reasoning settings remain limited, Meta's full safety report is still pending, and Wonder Full has no API. The biggest unknown is the memory feature: it will only prove itself once real people trust it with their inboxes, not demo data.

Disclosure: figures come from Meta's launch materials and third-party reporting as of October 13, 2026, and may change as Meta and independent evaluators publish more results.

Sources

Frequently asked questions

#ai#muse-wonder#meta#meta-superintelligence-labs#muse-code

More to read