Meta Releases Muse Wonder, Its First Frontier-Class AI Model

What happened

On Monday 12 October 2026, Meta released two new AI models from its Meta Superintelligence Labs team. According to Meta's launch post, Muse Wonder and Muse Wonder Full are its most capable models to date, and they are built for personal and professional agent tasks. Meta says the models were trained at its Prometheus data center in Ohio, in the United States. The company describes them as its first frontier-class models.

Access is split. Muse Wonder is available today to developers through the Meta Model API and through Muse Code, Meta's coding tool. Muse Wonder Full is available today to Muse Code subscribers only. Meta says Muse Wonder will also power experiences in its Muse personal agent app, starting with subscribers.

Pricing for Muse Wonder comes from TechCrunch's launch report. Muse Wonder Full has no per-token price because it is sold through a subscription. Per The Rundown's guide, it is open to eligible Muse Code Power and Maximum plans, where available.

Why it matters for businesses that automate work

Until now, most of the frontier models that businesses could call through an API came from OpenAI and Anthropic. Meta's entry adds a third option at the top tier. Meta's own post says the models are designed to handle tasks such as navigating an insurance claim, reading a lease, and building a working app. Those are the kinds of document-heavy, multi-step jobs that sit inside back-office work.

The launch also makes memory a central feature. Meta describes Active Memory, which carries context across conversations and brings it back when relevant, and Active Focus, which keeps the model working through blockers. For automation, the practical question is whether an agent can keep track of a long case without a person re-entering the details. Meta says Muse Wonder scores higher than rival models on its MemoryBench test. Meta reports this from its own testing.

An independent check exists, but it is narrow. Artificial Analysis scored Wonder Full at 56 on its Intelligence Index, up from 48 for Meta's earlier Muse Spark model. That places it level with Claude Sonnet 5.5 on the index and ahead of Claude Opus 5.5 on coding agents, according to the same report.

What changes in practice for a business that automates work

The change is less about a new tool and more about a new option for the model layer behind your automations. A company that already runs document handling or reconciliation through an API can add Muse Wonder as a second model and compare results on its own files. The price is set per token, so cost depends on how much text each job sends and returns.

Consider two examples. A freight forwarder in Rotterdam could test Muse Wonder on the bills of lading that its staff currently retype into an accounting system. A clinic group in Nairobi could test it on intake forms that need to be checked against a patient record. In both cases the test is the same: take 200 real documents, run them through the model, and count how many need a human to correct them. Neither example describes a result, only a test a business can run.

1. Pick one routine workflow with a clear right answer, such as matching invoices to purchase orders.
2. Run a sample of past cases through Muse Wonder via the API and record the error rate.
3. Keep a person on the exceptions and on any output that goes to a customer, a regulator, or a bank.
4. Check the data terms before sending personal or financial records to any outside model.

The limits are real. Muse Wonder Full cannot be called through the API, so a business cannot build it into its own systems at launch. Meta has not disclosed the parameter count for these models, according to ITdaily's review of the model card. Meta also says it added safeguards against misuse, including against sharing private information without permission. Those safeguards can also block legitimate requests, and a business should test where the refusals fall. Automation with a model like this still needs a person to own the process, check the output, and decide what happens when the model is wrong.

Automation is not the right move for every task. If a workflow runs only a few times a month, or if a single mistake would be costly and hard to reverse, a person working with a spreadsheet may be cheaper and safer. Frontier models are also more expensive per task than smaller models, so a simple classification job may not need one at all.

What to watch next

• Whether Muse Wonder reaches the consumer Muse app for all accounts, or only for subscribers.
• Whether Meta opens Muse Wonder Full to the API, which it has not promised.
• Independent benchmark results beyond Meta's own published tests.
• Any usage or subscription figures Meta gives with its next quarterly results, which analysts are watching as a test of return on its AI spending.