Meta's Muse Wonder Is Here, and It Costs $20 a Month

Meta released Muse Wonder on October 12, 2026, and called it the most advanced model it has. This is the first time Meta has built a model the company itself calls frontier class, meaning on par with the strongest models from OpenAI, Anthropic, and Google.

What sets this release apart is not its capability but its price. To use Muse Wonder through the Muse app, users must subscribe at $20 or $100 a month. The Muse app was previously free. Meta has long been known for releasing its models openly, and this release marks a change from that.

The model was built by Meta Superintelligence Labs, under the codename Watermelon. It was trained specifically for agents, programs that can carry out tasks on their own, such as booking a restaurant table or checking a schedule. According to Meta's AI chief Alexandr Wang, the problem with older agents was that they stalled when they hit an obstacle. Muse Wonder was trained to try other approaches instead of giving up and asking the user.

Two Models, Two Different Doors

Meta released not one model but two. Muse Wonder is available through the Meta Model API for paying developers. Muse Wonder Full, the stronger version, is only available to developers subscribed to Muse Code, Meta's coding agent, at $50 or $200 a month.

This means Meta's strongest variant is not reachable through an ordinary API. Developers who want it must enter the Muse Code ecosystem first. That is a business decision, not a technical one.

Wang said the Muse Wonder API price is higher than the standard subscription, because the model is much larger. But he claims it is still far cheaper than the leading models from OpenAI and Anthropic. The exact figure was not given in that interview.

The Overlooked Part: Memory

Muse Wonder was built with an emphasis on long-term memory. That does not sound dramatic, but it is what decides whether an agent is useful or annoying.

An agent with no memory will ask the same thing again and again. An agent with memory can recall that you prefer morning schedules, or that you already tried a certain restaurant. That ability is what separates an assistant from a plain chatbot.

It is also why Meta is pushing its model toward agents. If a model only answers questions, memory barely matters. If a model runs repeated tasks for you, memory is everything.

The Muse App Came First

Muse Wonder was not born in a vacuum. The Muse app shipped earlier, and shot straight to the top of Apple's App Store after it was introduced last month. The app is free, with subscription tiers for heavier use. The new model is a paid layer on top of it.

But the app also drew complaints. Some users said on social media that their agent accessed private messages without permission. Meta disputed that claim. Wang acknowledged there were cases where users granted permission early in a conversation and then forgot, and said the company is now educating users about it. Those complaints are worth noting, because they touch the same issue as the Muse Wonder release: how far an agent may act on a user's behalf.

That mix of popularity and complaints explains Meta's position. It has a large user base, and it needs to prove its agents can be trusted, not just downloaded in volume.

Why Meta Needs Money from AI

Meta has long made almost all its revenue from ads. For years the company tried to sell something else, and most of those efforts failed. Releasing Muse Wonder with a monthly price is its latest attempt to change that.

The model matters to Meta for two reasons. First, running a large model is expensive, and ads cannot cover all of that cost forever. Second, Meta wants to prove to investors that AI can be its own revenue source, not just a tool that makes ads more effective.

If Muse Wonder sells subscriptions, that sets a precedent. If it fails, Meta will still have AI, but as an ad engine rather than a subscription product.

How This Model Stacks Up Against Rivals

A "frontier class" claim only means something with a comparison. In the same week, Meta and Google released large models almost together: Muse Wonder on October 12, Gemini 4 Argon globally on October 13. Both claim top capability, and both raised prices.

What separates the two approaches is the terrain. Google positions Gemini 4 Argon for professional work and cyber defense. Meta positions Muse Wonder for personal and professional agents, things like booking a table or managing a schedule. Both target different pockets of the same user base.

On price, Meta chose a "cheaper than the leader" strategy. If that claim holds, Meta is trying to win not with the best model, but with a good-enough model at a lower cost. That is a sensible strategy for a player arriving late.

The Economics Behind Paid Agents

Agents are expensive to run, and that explains why Meta dares to charge monthly. An agent does not answer once and stop. It reads pages, tries several steps, fails, and tries again. Every attempt calls the model, and every call spends tokens. One successful task can burn dozens of times more tokens than a single question.

A plain model is sold per token, so its cost grows with use. An agent, by contrast, fits a subscription far better, because users cannot predict how many tokens one task will consume. This is a structural reason, not just a marketing choice. When the cost per task is hard to estimate, a subscription moves that risk from the user to the company.

Meta took this position deliberately. It absorbs unpredictable costs and charges a fixed price in return. If the model is efficient enough, the gap becomes profit. If not, a $20 plan can be expensive for Meta itself.

What Is Still Unclear

Several things cannot be verified from this announcement.

• Official benchmark numbers. Meta has not published a full comparison of Muse Wonder against GPT-6 or Claude Opus 5.5. The "frontier class" claim remains a company claim until independent numbers exist.
• The exact API price. Meta only said "cheaper", with no figure.
• The performance of Muse Wonder Full. Because it is only reachable through Muse Code, independent testing will be harder.

Meta also claims Muse Wonder has its "strongest safeguards to date" against cyber misuse. But those protections currently exist only in the paid model. That means free users run a version without the full protection.

Analysis: The Shift from Open to Subscription

Meta's decision is interesting because it contradicts its old image. Llama, Meta's previous model family, was released with open weights. Anyone could download it, run it, and modify it. That is what made Meta popular with developers.

Muse Wonder moves the other way. Its weights are closed, access is locked behind a paid API, and its strongest variant is confined inside another product. This is not a betrayal; it is a business strategy. Three forces explain the turn.

• Revenue. Open models build an ecosystem but do not generate direct income. Subscription models do.
• Cost. Running a frontier model is expensive, and ad revenue cannot cover it without limit.
• Control. A closed model lets the maker set safeguards, pricing, and access on its own terms.

Open models are good for building an ecosystem but do not generate direct revenue. Closed subscription models do.

The question to watch: will Meta open Muse Wonder's weights later, as it often did with Llama? If yes, this is only a matter of timing. If not, it is the end of Meta's open-model era.

For users in Indonesia, the effect is direct. As long as Meta's models can be self-hosted through open weights, costs can be kept down and data does not leave your own server. If Muse Wonder stays closed, the only way to use it is through a paid API, which means a subscription cost that shifts with exchange rates and company policy.

The question that remains: if the best models are only reachable through a subscription, what does "open AI" mean for developers in countries without a large budget?