Meta launched Muse Wonder at $4 and $20 per million tokens

A team that was choosing between Opus 5.5 and GPT-6 Astra for an agent that has to remember a user's work across a long session now has a third option, priced in the same bracket. TechCrunch reports that Meta's Superintelligence Lab released Muse Wonder, the company's first frontier-class model, and a larger variant called Muse Wonder Full. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. Wonder Full stays exclusive to Muse Code subscribers and is not reachable through the API.

Meta describes Wonder as its competitor to Anthropic's and OpenAI's current frontier releases. Russell Brandom at TechCrunch quotes Meta's own benchmark claim that Wonder reaches "comparable benchmark scores to Fable 5.1 and GPT-6 Astra." The comparison is Meta's, run and reported by Meta on its own test suite, with no independent run alongside it.

What Meta says is new in the model

The capability Meta leads with is called Active Memory, which the company says lets the model "retain and synthesize information given by the user" across extended sessions rather than treating each request as a fresh context. On a benchmark Meta calls MemoryBench, which measures continual learning, Meta says Wonder scores higher than Fable 5.1, Opus 5.5, and GPT-6 Astra. That is another vendor-run comparison, and MemoryBench is a Meta-defined test rather than a widely cited public one.

TechCrunch names research scientist Kavya Kopparapu as the person who presents the memory feature in Meta's accompanying launch video. The models were trained at Meta's Prometheus data centre in Ohio, which the company said earlier in the year had come online.

Where this lands against the other frontier prices

The $4 input and $20 output rate puts Wonder in the same bracket as GPT-6 Sol and Opus 5.5 at list price, and above Google's $2 input and $10 output rate for Gemini 4 Argon, which went to general availability the same week. TechCrunch reads the pricing match with Sol and Opus as deliberate: Meta is positioning Wonder as a direct replacement rather than a cheaper option, and the memory claim is what the company offers in place of a lower price.

Wonder Full is the harder one to assess. Meta is holding it back from the API and keeping it inside Muse Code, its coding subscription, so a team outside that subscription cannot test the larger model against the one that is open to them. Meta's decision with Wonder Full matches a pattern other labs have used: the best model stays behind the first-party product, and the API gets the second-best. A team that was planning to pay for the larger model as an API call for a one-off job is the one that cannot get it.

For a team already running on Opus 5.5 or GPT-6 Sol, Wonder asks for a comparison on a real workload rather than on a benchmark score. The question is whether Active Memory survives the kind of multi-turn agent run the team actually ships, or whether it only shows itself on tests Meta designed for it. On that question the launch materials cannot help: a Meta-run score on a Meta-defined benchmark is marketing, not an answer.

Source

Meta releases its first frontier model, Muse Wonder, by Russell Brandom in TechCrunch.