Menlo Park, Calif. — Meta Releases Its First Frontier AI Model

Meta's Superintelligence Lab released two new large language models on Monday — Muse Wonder and Muse Wonder Full — marking the company's first attempt at a frontier-class system, according to TechCrunch.

The release follows Muse Spark, a lighter model Meta built for low-latency tasks. The new pair is designed for scale rather than speed: Meta says the larger models post higher benchmark scores and handle complex work, including coding, better than Spark did.

What did Meta actually release?

Two models make up Monday's launch. Muse Wonder is available through the Meta Model API. Muse Wonder Full is restricted to subscribers of Muse Code and will not be offered through the API at all, TechCrunch reports. Both versions will also run the Muse agent, Meta's consumer assistant product.

Meta's own blog post, quoted by TechCrunch, frames the release in terms of the company's broader AI roadmap: "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence."

How much does it cost to use?

Meta set API pricing for the standard model at $4.00 per million input tokens and $20.00 per million output tokens, according to the company's published rates cited by TechCrunch. That puts the model in roughly the same price tier as GPT-6 Sol and Opus 5.5 — both competitor offerings referenced in the same report.

By the numbers

• $4.00 — cost per million input tokens for Muse Wonder via the Meta Model API
• $20.00 — cost per million output tokens for the same model

Wonder Full carries no public API price because it is bundled into a Muse Code subscription rather than sold by the token.

How does Wonder compare to GPT-6 Astra and Fable 5.1?

Meta's model card, described in the TechCrunch piece, shows the new release reaching benchmark scores comparable to Fable 5.1 and GPT-6 Astra — two systems built by rival labs. TechCrunch does not publish the underlying numeric scores from that card, so no specific benchmark figures beyond the pricing above are independently verified here.

On one metric Meta does highlight, the company says its release outperforms the field. The firm points to MemoryBench, a benchmark built to test continual learning and memory construction, where Meta's testing has the new model scoring higher than Fable, Opus and Astra, per the TechCrunch report.

What is "active memory" and why does Meta keep mentioning it?

Meta calls the new system's recall capability "active memory," shorthand for a model's ability to retain and reuse information a user has supplied across sessions. The company says this was a deliberate design target, not a side effect of scaling up the parameter count.

"We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months and beyond," research scientist Kavya Kopparapu said in a launch video cited by TechCrunch.

The emphasis tracks with Meta's stated ambition for the Muse agent: a personal assistant meant to carry context forward over long stretches of use rather than starting fresh with every prompt.

Where was the model trained, and what comes next for Meta's AI lab?

TechCrunch reports the models were trained at Meta's Prometheus data center in central Ohio, a facility that came online earlier this year. The release is the first frontier-tier launch since Scale AI founder Alexandr Wang restructured Meta's AI division in August of last year. In the period between that reorganization and Monday's announcement, Meta shipped the lightweight Spark model and the Muse agent, but had not matched the state-of-the-art releases coming from OpenAI and Anthropic — until now.

Whether Wonder and Wonder Full close that gap on independent benchmarks, rather than Meta's own model card, is not yet established in the public record. TechCrunch's report draws its comparative figures from Meta's internal testing; no third-party evaluation of the new models was cited in that story.

For readers tracking the token economics, the $4-per-million-input and $20-per-million-output pricing is the clearest external data point available so far, since it was set by Meta rather than estimated from benchmark claims.