Muse Wonder Explained: Meta’s First Frontier AI Model Costs $4 Per Million Tokens

Meta spent 14 months rebuilding its AI operation, and on October 12 it showed the first big result. Muse Wonder and Muse Wonder Full are the company’s first attempt at a frontier model, released by Meta Superintelligence Labs, TechCrunch reported.

Alexandr Wang, who leads Meta’s AI efforts, announced the launch on X. Here’s his post:

What Muse Wonder actually is

Muse Wonder is the bigger sibling of Muse Spark. Spark was built to be lightweight and quick. Wonder is larger, and Meta says it’s built for personal and professional agentic tasks.

Agentic just means the AI does the work instead of only answering questions. Meta’s examples include working through an insurance claim, making sense of a lease, and building a working app.

Two features carry most of the pitch. Active Memory lets the model carry details across conversations. Active Focus keeps it going when a task gets hard, so it tries other approaches, recovers from errors, and checks in before anything important.

On the technical side, OpenRouter’s listing shows a 1 million token context window and support for text, images, video, audio and PDFs. Meta trained the models at its Prometheus data center in Ohio.

Read More: Meta Muse Charm: The Keychain AI Gadget That Could Change How India Uses Apps

Two models, two doors

Here’s where it gets awkward for anyone hoping to try the best version. Muse Wonder is available through the Meta Model API and Muse Code, and it’s headed to the Muse app soon, starting with subscribers.

Muse Wonder Full, the stronger one, is only for Muse Code subscribers. It isn’t served through the API at all.

CNBC reported what entry costs. Muse app users need a $20 or $100 monthly plan to get Wonder, while Wonder Full sits on Muse Code’s $50 or $200 plans. Inside Meta the model went by the codename Watermelon, a joke Wang kept alive with a fruit emoji in his launch post.

The $4 question

On the API, Wonder costs $4 per million input tokens and $20 per million output tokens. TechCrunch puts that roughly level with GPT-6 Sol or Opus 5.5.

It’s also a big jump from Meta’s own earlier model. OpenRouter lists Muse Spark 1.3 at $1.25 and $4.25 for the same units, so Wonder costs about three times as much on input and nearly five times as much on output.

Wang told CNBC the pricing would stay “meaningfully below” the leading models from OpenAI and Anthropic. That’s hard to square with a price sheet that matches Opus 5.5, and nothing we found from Meta explains the gap.

Meta’s scorecard, and why to hold the applause

In Meta’s own testing, Wonder and Wonder Full score comparably to Fable 5.1 and GPT-6 Astra. Meta also says Wonder beats Fable, Opus and Astra on MemoryBench, a test of how well models remember. Wang went bigger on X, calling it the best model in the world for personal agents and professional agents.

Those are Meta’s numbers. Interesting Engineering noted that the announcement gave no numerical scores in the material it reviewed, which makes any gain hard to size. Meta says a full safety report is coming soon.

What this means for Indian founders

Meta’s announcement doesn’t spell out regional availability, and the Muse app launched in the US first, on September 8. Indian users shouldn’t count on the consumer app showing up soon.

The API is a different story. It’s a developer product priced in dollars, and Indian startups building agents for support, sales or back-office work already weigh model bills against thin margins. Wonder gives them one more option to test, not a reason to switch.

Treat it like a D2C brand treats a new ad channel. Test small, measure cost per outcome, then decide whether to scale. It also keeps the pressure on homegrown model builders such as Sarvam AI to show where they win, whether that’s price, language or fit for Indian use cases.

Read More: Did Alexandr Wang’s Meta Muse Justify a $14.3 Billion Bet? It Beat ChatGPT, Then the Stock Slipped

The real test is who pays

CNBC framed the launch as a test of whether people will pay for AI, and that’s the sharper reading. The Muse app is free, with subscriptions for heavier use, and Meta wants AI revenue beyond its ad business. Wonder puts its strongest models behind paid plans.

It’s also the first frontier-class release since Wang restructured Meta’s AI effort in August 2025, per TechCrunch. Free got Meta users in the door. A $20 to $200 price will show whether they stay.

Our take: if Wonder Full justifies $50 to $200 a month for coders, Meta has a second business. If it doesn’t, benchmark charts won’t rescue it.

Before you switch anything

• Run Wonder on your own agent tasks through the API, and track cost per finished task, not just cost per token.
• Don’t plan around Wonder Full yet, since Meta hasn’t said whether it will ever reach the API.
• Wait for Meta’s safety report and independent benchmarks before putting Wonder near customer data.

Building on AI agents? Tell us in the comments which model you’d test first, and share this with your tech lead. For more on Meta’s AI push and what it means for Indian startups, keep exploring StartupIndiaX.

FAQs