Meta Platforms released its first frontier-class artificial intelligence model on October 12 and put a price on it, ending the company’s run as the industry’s most generous donor of free models.
Muse Wonder, built by Meta Superintelligence Labs, is designed for agents that do work on someone’s behalf rather than answer questions in a chat window. The company’s examples are ordinary and specific: working through an insurance claim, making sense of a lease, tracking where a household’s money went.
Developers reach the standard model through the Meta Model API and a coding product called Muse Code. Meta set the price at $4 per million input tokens, $20 per million output tokens and $0.20 for cached input, according to the company. Muse Code subscriptions run from $5 a month to $200. The stronger variant, Muse Wonder Full, comes only with a paid plan and is not sold through the API.
The rate card is the news. Fourteen months ago Meta pitched open weights as its route into the market, publishing models that anyone could run on their own servers. This release goes the other way. It also arrives while rivals cut prices: Google opened Gemini 4 Argon to all developers on October 13 with a limited-time rate of $2 per million input tokens and $10 per million output tokens, below what Meta is asking.
There is a commercial reason to charge. Advertising still pays Meta’s bills, and investors have spent three years asking when AI spending turns into an AI revenue line. The Muse app, released last month, rose to the top of Apple’s App Store, and subscriptions to it start at $20 a month and reach $100, according to CNBC. Users of the app need one of those plans to reach the new model.
The benchmark scores say the model is close to the front. Artificial Analysis, a third-party evaluator, scored Muse Wonder at 56, one point behind Anthropic’s Claude Opus 5.5 and above OpenAI’s GPT-6 Astra, Google’s Gemini 4 Argon and Anthropic’s Claude Fable 5.1, which it placed at 53. Meta says the model passed Opus 5.5 on tests of long-horizon coding and memory.
Meta trained the family for work that unfolds over time. The company describes two features it built for that purpose: Active Memory, which carries context across conversations and brings back details a task needs, and Active Focus, which lets the model try another approach when it hits a wall, recover from its own errors and check in with the user before doing anything consequential. The company calls those traits the difference between an assistant that stalls and an agent that finishes.
The safety design follows the pricing. Meta’s blog post says the models carry “our strongest safeguards to date” against cyber misuse, and those protections come with the paid model first. Muse Wonder Full is trained to refuse requests for harmful acts and to withhold private information without permission, the company said.
Alexandr Wang, Meta’s chief AI officer, told CNBC in an interview on Monday that the goal was to keep agents from stumbling on tasks like finding an open dinner reservation. “When you do real-world tasks, you’re going to get roadblocks and you’re going to get things that come up or various setbacks,” he said. “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.”
Mark Zuckerberg said the company has already begun work on more advanced products and that model development “may notably accelerate from now on,” according to a person close to the company.
The launch lands in a market that is still learning how to charge for agentic software. Chatbots are billed by the token, which maps loosely onto how much text goes in and comes out. An agent is different: it loops, retries, reads files and calls tools, so a single finished task can consume millions of tokens whose cost the customer cannot easily predict. Per-token pricing punishes the customer for the model’s inefficiency, the argument runs, which is why several software vendors have started quoting per-task or per-outcome rates instead.
Meta is choosing the meter anyway, at a moment when its costs are rising. Training a frontier model requires compute that Meta buys at market prices, and the company’s capital spending has climbed alongside everyone else’s. Charging for access at least gives shareholders a number to compare against the budget, even if the number is small at first.
The open-weight decision carries its own risk. Meta built a developer following by letting people download its models, and that following has alternatives. Startups that depend on a hosted model will test Muse Wonder against cheaper options from Google and from China’s open-weight labs, and the price list gives them a reason to look. A developer who wanted Meta’s weights at no cost now has to decide whether the performance gap justifies a subscription.
What Meta gets in return is a clearer business. Subscriptions are recurring, they are measurable, and they do not depend on persuading advertisers to spend on new formats. The company has told investors for two years that AI would produce revenue of its own. With Muse Wonder, the claim moves from a slide into a price list.
Whether users pay is the test that starts now. Meta’s results over the next two quarters will show how many developers sign up for Muse Code, and how many app users accept $20 a month for an agent that plans their day. Either way, the era of Meta giving away its best model is over, and the company has told the market what it thinks the work is worth.
