Skip to content

Meta Releases Muse Wonder, Its First Frontier Model: By Its Own Tests It's Neck and Neck with Rivals, and Its Biggest Boast Is That It'll Remember You for Months

1 min read
Share
Meta Releases Muse Wonder, Its First Frontier Model: By Its Own Tests It's Neck and Neck with Rivals, and Its Biggest Boast Is That It'll Remember You for Months

Meta's superintelligence lab released two new models on Monday, Muse Wonder and Muse Wonder Full. The company is presenting them as its first attempt at a frontier model - the space where OpenAI and Anthropic have been competing while Meta watched from the stands.

Until now Meta has played it safe. Muse Spark was a lightweight model, built to answer quickly rather than to beat the best. The Wonder models are larger and, according to tests published in the model's technical report, achieve results comparable to Fable 5.1 and GPT-6 Astra. The key part of that sentence is "according to Meta's tests": the numbers were measured by the company selling the model.

The price shows how seriously Meta is taking this. Through the Meta Model API, Wonder will cost 4 dollars per million input tokens and 20 dollars per million output tokens - roughly the same as GPT-6 Sol or Opus 5.5. The more powerful version, Wonder Full, won't be available via the API at all: only Muse Code subscribers will get it, and both models will also power the Muse agent.

The biggest emphasis isn't on coding, but on memory. Meta calls it "active memory" - the model's ability to retain and connect everything the user has told it. On MemoryBench, a test measuring continual learning and memory building, Wonder outperformed Fable, Opus and Astra, according to the company. "We put enormous effort into making the model remember what matters to you, for days, weeks, months and longer," researcher Kavya Kopparapu says in the video presenting the models.

This is worth pausing on. A company whose business has for years rested on knowing as much as possible about you in order to sell you ads is now building an assistant whose main selling point is that it never forgets. Meta calls the goal "personal superintelligence". Is that just a politer name for a more detailed dossier?

The models were trained at the Prometheus data centre in central Ohio, which came online this year. It's the first frontier release since Scale AI founder Alexandr Wang reorganised Meta's AI efforts in August last year. Fourteen months later, Meta has finally entered the race. It remains to be seen whether anyone outside the company will confirm the numbers - and how many people actually want an assistant that remembers them for months.