Muse Wonder: what Meta's new model changes when you work alone

Every Monday, you paste the same paragraph into your AI assistant: who your client is, what you promised them, the tone of your quotes. And every Monday, it has forgotten everything.

Muse Wonder is the new AI model from Meta, announced on 12 October 2026 alongside a more powerful version, Muse Wonder Full. Meta trained both to remember you from one conversation to the next, and to see a task through to the end.

At launch, Muse Wonder was pay-as-you-go, and Muse Wonder Full was subscription-only. Meta's consumer app, meanwhile, was not open in Europe.

So let's look at what it can do, what it costs and what happens to your data, even if you have never written a line of code.

What is Muse Wonder?

It comes from Meta Superintelligence Labs, Meta's AI lab led by Alexandr Wang. It is Meta's first model aimed at the level of the best models from OpenAI and Anthropic.

According to Meta's announcement, it is built for AI agents, the assistants that act on your behalf inside your tools. Meta's examples are very down to earth: handling an insurance claim, understanding a lease, building the app for a business that is just starting.

You cannot download it, unlike Meta's older Llama models. To use it, you always go through Meta's servers.

What can Muse Wonder do that others can't?

Meta highlights two abilities, with names of its own:

• Active Memory, the memory, which keeps what matters to you from one conversation to the next and brings it back when the task needs it;
• Active Focus, the persistence, which makes it try another approach when a task gets stuck, ask a question if needed and get your approval before anything important.

Picture an assistant that keeps a notebook for each client. No more telling it again that Smith pays at 60 days and hates calls on Fridays.

For a long time I kept a "client context" file open at all times, just to paste it into every new conversation (yes, every single morning 🥲). That is exactly the habit Meta wants to get rid of.

The announcement does not say what the model keeps about you, or where it stores it. The full safety report was not out on launch day, and Meta promised it for later.

Is it as good as Claude and GPT-6?

On memory tests, Meta says it beats Claude Opus, Claude Fable and GPT-6 Astra. Those are Meta's own tests.

The most useful ranking comes from Artificial Analysis, an independent company that scores models. Its index combines ten evaluations (reasoning, knowledge, coding, professional tasks), and the higher the score, the better.

Here are the scores published at launch:

Muse Wonder Full matches Claude Sonnet 5.5, without catching Opus 5.5. The version you can plug into your own tools, Muse Wonder, sits four points lower.

How much does Muse Wonder cost?

Through the API, Meta's pricing page lists $4 per million tokens of input and $20 of output at launch. Input is what you send. Output is what it writes back.

Take a reply to a client's email, roughly 2,000 tokens to read and 1,000 to write. It costs just under 3 cents, so $28 for a thousand emails.

Here is the same job with the models we have already compared:

Muse Wonder therefore costs twice as much as Sonnet 5.5, for a lower score. It is priced like Claude Opus 5.5, without its level.

Muse Wonder Full, on the other hand, is not sold pay-as-you-go. You need a Muse Code subscription, Meta's answer to Claude Code. Two plans included it at launch:

• Power Usage, at $50 a month, with twenty times the usage of the basic plan;
• Maximum Usage, at $200 a month, with eighty times that usage.

For the Muse app, CNBC reports subscriptions at $20 and $100 a month to get Muse Wonder. Assuming you can install it at all.

Can you use Muse Wonder from Europe?

For the app, no. Muse launched in the United States on 8 September 2026, then in Canada on 18 September. Meta gave no date for Europe.

It would not be the first time. Meta AI, its chat assistant, took eighteen months to cross the Atlantic (September 2023 in the US, March 2025 in Europe).

For the API, things are less clear. Meta's geographic use policy, updated on 12 October 2026, does not exclude Europe. Its restricted territories include Russia, China and Iran.

Meta still says the API is not available everywhere and that it decides country by country, without publishing a list. The simplest way to find out is to open a developer account and see whether yours goes through.

What happens to the data you send it?

The question matters as soon as you work for clients. Meta offers two modes:

• Standard mode, where Meta does not use what you send to train its models. The pay-as-you-go API only runs in this mode;
• Contributor mode, where your prompts, your code, your files and the replies are used to improve future models. It is the default setting for Muse Code subscriptions.

Then comes the GDPR, the European rules on personal data. Technspire, a Swedish IT services company, went through the API terms at the end of September.

It found no data processing agreement and no European hosting. Until Meta signs one, do not send your clients' personal data. A public text or a product page, however, is no problem.

On the app side, CNBC also reports that some American users accused the agent of reading personal messages without their consent. Meta disputes this, saying those people had given consent earlier and forgotten about it.

Should you switch to Muse Wonder?

It depends on what you hand it:

• You code projects alone that run for days. Long coding sessions are where Meta claims its biggest lead, and the $50 plan is worth a try, with Contributor mode off;
• You automate emails or documents, in n8n for example. Sonnet 5.5 or GPT-6 Sol do the job for half the price, and our n8n course shows how to switch models at a single step;
• You handle client data. Wait until Meta publishes a data processing agreement and European hosting;
• You use an assistant to write and think. There is nothing to do, since the Muse app was not open in Europe at launch.

If you are torn between the two cheaper models, we put them head to head in our comparison of GPT-6 Sol and Claude Sonnet 5.5. And if price comes first, Mistral Large 4 does the same job for $7.

Plugging Muse Wonder into your tools (for developers)

You can skip this part if you don't code.

In the API, the model is called muse-wonder-1.0. According to Meta's documentation, it reads text, images, audio, video and PDFs, and replies in text.

Its context window reaches about one million tokens, with no surcharge for long requests. You can send it an entire client file instead of sorting the documents by hand.

The API accepts the same formats as OpenAI's and Anthropic's. An existing script can be tested by changing the address and the model name, within a limit of 3,000 requests per minute.

Don't worry, memory is not Meta's exclusive: you can prepare it yourself. Keep your "client context" in a clean document, it will serve you whatever the tool! 😉

Frequently asked questions about Muse Wonder

Muse Wonder is an AI model from Meta, announced on 12 October 2026 to power agents that remember you from one conversation to the next.

It comes in two versions. Muse Wonder is used through Meta's API, and Muse Wonder Full through a Muse Code subscription.

At launch, $4 per million input tokens and $20 per million output tokens through Meta's API, the same price as Claude Opus 5.5. A thousand replies to client emails come to about $28.

Muse Wonder Full comes with a Muse Code subscription, from $50 a month.

The Muse app was not open in Europe at launch, and Meta had given no date.

For the API, Meta's geographic policy does not exclude Europe, but Meta opens access country by country without publishing a list. You need to open a developer account to find out whether yours is supported.

Through the pay-as-you-go API, no: Meta does not train its models on what you send in that mode.

Muse Code subscriptions turn on Contributor mode by default, which does use it. Switch it off before sending a client's code, because the change only applies from then on.