Meta Muse Wonder is here: a frontier AI model for agents, and it is not free
Meta launched Muse Wonder and Muse Wonder Full on October 12, 2026. Artificial Analysis scores it 52, the API costs 4 and 20 dollars per million tokens, and the best tier sits behind Muse Code plans.
For most of the last two years, Meta was the company that gave AI away. Open weights, free chat in WhatsApp and Instagram, and a lot of talk about putting a model in every pocket. On Monday, October 12, 2026, Meta did something that feels very different. It launched its most capable models so far, Muse Wonder and Muse Wonder Full, and put both of them behind a paywall.
Meta describes the pair as frontier models for personal and professional agents. That means software that does not only answer a question, but plans a run, reads a lease, fills in an insurance claim or builds a small app for you. Meta says the models were trained at its gigawatt scale Prometheus data center and that they are the result of a ground up overhaul of its AI work that started fourteen months ago.
The launch landed with real weight. Reuters reported that Meta shares rose about 4 percent in afternoon trading. CNBC interviewed Meta AI chief Alexandr Wang, who said the model was developed under the codename Watermelon. The Verge noted that Meta claims its new models can go toe to toe with GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, based on its own internal testing. Here is what is actually new, what the independent numbers say, and where the catch is.
The three things you need to know
- What is new. Meta released Muse Wonder and the stronger Muse Wonder Full. Both are built for agents that keep context about you across conversations and keep working when a task hits a roadblock. Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.
- The numbers. Independent testers at Artificial Analysis give Muse Wonder at its highest reasoning setting a score of 52 on their Intelligence Index. The API price is $4 per million input tokens and $20 per million output tokens, with a context window of about one million tokens.
- The catch. It is not free. According to CNBC, people who want Muse Wonder inside the Muse app will need a $20 or $100 monthly plan, and Muse Wonder Full is only available to developers on Muse Code plans of $50 or $200 a month.
What Meta actually launched
Meta is splitting this release into two tiers. Muse Wonder is the general model. Developers can call it through the Meta Model API today, and it is also available inside Muse Code, Meta's coding agent. Meta says Muse Wonder will soon power its Muse personal agent app as well, starting with subscribers, and that it will later use Muse Wonder to build specialized models that can reach everyone who uses Muse. Meta has not shared a timeline for either step.
Muse Wonder Full is the bigger sibling. For now it only lives inside Muse Code. That is a clear signal of where Meta thinks the most valuable work is: long software projects where a model has to keep track of requirements, edit a code base and survive many rounds of testing.
Meta frames both models around two ideas. The first is Active Memory. In plain words, the model should remember what matters about you across conversations and pull up the right detail when it becomes relevant. The second is Active Focus. The model should keep pushing toward a goal, try a different approach when something breaks, ask for clarification when it needs to, and check in before it takes an important action.
Wang explained the second idea to CNBC with a very ordinary example. Real tasks, like booking a dinner reservation and checking availability, run into setbacks. Meta says it trained the model to try other approaches instead of stalling, and to check back with the user when needed. Anyone who has watched an AI agent give up after the first error message will understand why that matters.
Meta also points to domain work. Its research post mentions evaluations in law, finance and health, including Harvey's Legal Agent Benchmark, which asks models to work with source documents and produce legal work products. For coding and long tasks, Meta names FocusBench and SWE-Marathon, a benchmark for long running software projects. The launch page shows demos such as a playable 3D kitchen game built from images, a running plan made in a browser, custom merchandise ordered from photos on an Instagram profile, and a wedding seating chart with complex matching rules.
What the independent numbers say
Company benchmark tables are always a starting point, not the final word. That is why the Artificial Analysis page is the most useful number in this story. Artificial Analysis runs its own suite of ten evaluations, including Terminal-Bench, SciCode, Humanity's Last Exam and several agentic work tests, and combines them into one Intelligence Index.
On that index, Muse Wonder at the Max setting scores 52. Artificial Analysis lists that as well above average among comparable reasoning models in a similar price tier, where the median is 27. In its intelligence ranking the model sits at number 9 out of 219 models in that comparison class. That is a strong debut, but it is not a clean sweep of the top spot.
The same page shows the cost side. Input tokens cost $4 per million and output tokens cost $20 per million on Meta's API, with a 95 percent discount for cached input. Artificial Analysis calls that somewhat expensive for input and expensive for output compared with the median of $2 and $10. The model accepts text, image, speech and video input, produces text output and has a context window of about one million tokens.
There is one more detail that matters for anyone who pays per token. Muse Wonder is chatty. To run the full Intelligence Index it generated 130 million output tokens, against a median of 85 million for similar models. Artificial Analysis puts the average cost at $4.52 per index task. It also measured about 83 output tokens per second and a time to first token of roughly 19 seconds, which is slower than average. For long agent runs that work for minutes, speed may matter less than reliability. For quick chat, it will feel slow.
The paywall is the real story
The most important part of this launch might not be the score. It is the business model. CNBC framed the release as a test of whether people are willing to pay Meta for AI. Users of the Muse personal agent app will need a monthly subscription of $20 or $100 to get Muse Wonder. Developers who want Muse Wonder Full need Muse Code at $50 or $200 a month.
On the API side, Wang told CNBC that Muse Wonder will cost more than the standard subscriptions because it is a much larger, frontier level model, but that it is still priced meaningfully below the leading models from OpenAI and Anthropic. That fits Meta's broader goal of building AI revenue that does not depend only on ads.
There is also a privacy angle that should not be ignored. CNBC reported that some Muse users complained on social media that the agent accessed personal messages without authorization, which Meta disputes. Wang said some users may have authorized communications earlier in a conversation and then forgotten. When the whole pitch of a model is that it remembers you and acts for you, trust is not a side topic. It is the product.
Safety claims and what is still missing
Meta says Muse Wonder offers its strongest safeguards to date against chemical, biological, radiological and nuclear misuse and against cyber misuse, while keeping frontier cyber capabilities available for defensive use under its Meta Superintelligence Scaling Framework. Meta also calls it its most aligned model so far.
Those are claims, not test results. Meta says a full report with additional safety evaluations will be available soon. Until that report is public, outsiders cannot check how the safeguards were measured. CNBC also noted that the new protections come with the paid model for now. Given the week we just had, with AI agents causing real world incidents and governments asking for incident reporting, that report deserves close reading when it arrives.
How it compares
Meta's own comparison puts Muse Wonder and Muse Wonder Full next to GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The Verge pointed out that it is still unclear how the models stack up against Google's Gemini 4 Argon, whose capabilities are still coming into focus. Independent results like the Artificial Analysis index will tell us more over the next weeks, once more testers have real API access and run their own workloads.
What we can say today is simple. Meta is back in the frontier conversation with a model that independent testers rank near the top, at a price that is higher than the median but, according to Meta, below its biggest rivals. And it uses a lot of tokens to get there.
Why it matters
For two years, the frontier race looked like a contest between OpenAI, Anthropic and Google, with Meta playing a different game through open weights and free products. Muse Wonder changes that picture. Meta now sells a closed frontier model, charges for access and targets the same agent and coding market as everyone else.
That matters for three groups. Developers get another serious option with a one million token context window and a public API. Consumers get a personal agent that wants to remember more about them, which raises the bar for privacy controls. And the market gets a real test of whether people pay Meta directly for AI instead of paying with their attention.
Dany's take
I like that Meta finally shows up with a model that independent testers take seriously. A score of 52 on the Artificial Analysis index is not marketing, it is a measured result. But I would not switch everything over tonight. The model is slow to start, it burns a lot of tokens, and the most important safety report is still marked as coming soon. The part I find most interesting is the paywall. Meta spent years telling us AI should be free and open. Now its best model costs money and lives inside its own apps. That is not good or bad by itself, but it means we should judge Meta like every other paid AI vendor: on reliability, privacy and honest numbers. Would you pay for Meta's AI agent? I am curious what you think.
Sources
Source: research.meta.ai