Meta Muse Wonder: Pricing, Access and What It Means for You

On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable AI models. Muse Wonder is already available through an API at $4 per million input tokens and $20 per million output tokens. It matters to anyone running AI on a website or in a business, and to anyone who wants AI agents to be able to use their site.

What we know

What Meta launched: two models, Muse Wonder and Muse Wonder Full, announced on October 12 on Meta's research blog as its "most capable" models, built for personal and professional agents.
Where you can use them: Muse Wonder is available from launch day in the Meta Model API and in Muse Code, Meta's coding tool. Muse Wonder Full is only available with a Muse Code subscription.
When it reaches the Muse agent: Meta says Muse Wonder will "soon" power its Muse agent, starting with subscribers. No date was given.
API pricing: according to the model page on Meta's developer portal, muse-wonder-1.0 costs $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens, with a 1-million-token context window.
Wonder Full pricing: there is no per-token API price. It is included in Muse Code's Power Usage ($50 a month) and Maximum Usage ($200 a month) plans, according to the same page.
What Meta promises: coding, agents that work for hours, and "advanced computer use": agents that see and operate browsers and apps, per Meta's model page.
The comparisons: Meta publishes a table where its models compete with GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. These are Meta's own tests. The Verge notes it is unclear how they compare with Google's Gemini 4 Argon.
Safety: Meta says a full report with more safety evaluations will be published "soon."

What changes and what doesn't

The menu changes. There is one more frontier model you can buy by API, and Meta's page says the API works with OpenAI SDK-compatible clients. Trying it in a tool that already speaks that format should mostly mean changing the endpoint and the key.

The cheap option doesn't change. Muse Spark 1.3, Meta's previous model, still costs $1.25 input and $4.25 output per million tokens. Wonder is about three times pricier on input and almost five times on output.

Nor has it been proven outside Meta. The performance figures come from the company, and the safety report is not out yet.

A worked example, calculated for this article: a website assistant handling 1,000 conversations a day, with 3,000 input tokens and 500 output tokens each, would cost about $22 a day on Muse Wonder (about $660 every 30 days). On Muse Spark 1.3 it would be about $5.90 a day (about $176 every 30 days). If much of the input repeats in every conversation and hits the cache, Wonder's cost drops.

How to tell if it affects you

If you run AI on your site, list what your current model actually does. For FAQs or sorting messages, a cheap model is usually enough. Wonder makes sense for long tasks: reviewing documents, coding or chaining several steps.
Do your own math. Multiply daily conversations by input and output tokens per conversation and apply the prices on Meta's page. Test with 50 to 100 real cases before changing anything in production.
Watch the cheap tier. The Muse Spark 1.3 page lists a "contributor" option at $0.10 input, labeled as used to improve Meta's products. Don't send it customer data or internal information.
Get your site ready for agents. If agents are going to drive browsers, your site has to be usable without guesswork: buttons with clear text, labeled form fields, and prices, hours and terms in text, not only in images.
Make sure key content loads plainly. If it only appears after several clicks or pop-ups, an agent will struggle to find it, just like a search engine.

Sources

Updates: this article will be updated when Meta publishes its safety report or brings Muse Wonder to the Muse agent.