56 points on the Artificial Analysis Intelligence Index — that's where Meta's newest model landed at maximum reasoning effort. It's the second-best score of any AI system on the planet, ahead of GPT and Google's models, trailing only Anthropic's Claude Opus 5.5.
On October 12, Meta shipped Muse Wonder and Muse Wonder Full, the first models the company itself is willing to call "frontier" — meaning built to go toe-to-toe with OpenAI, Anthropic, and Google rather than just catch up. The standard Wonder is live on the Meta Model API; the beefier Full version is reserved for Muse Code subscribers. Pricing isn't cheap: $4 per million input tokens and $20 per million output tokens, roughly in line with GPT-6 Sol or Claude Opus 5.5.
The headline feature is "active memory" — the model holds on to details across days, weeks, and months instead of forgetting them the moment a chat ends. Meta research scientist Kavya Kopparapu put it plainly: "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." A companion feature, "active focus," keeps the model working through a task after a mistake instead of giving up, asking clarifying questions along the way. Meta is pitching the pair for legal, financial, and medical work, plus coding.
That's a reversal for a company whose Llama models had fallen visibly behind GPT and Claude, and whose lightweight Muse Spark release earlier this year didn't change the conversation. The turnaround traces back to August 2025, when Scale AI founder Alexandr Wang took over the lab and rebuilt the team from scratch. Wonder was trained at the Prometheus data center in Ohio, which only came online earlier this year.
It's the first time Meta's numbers, not just its budget, put it in the same sentence as the frontier leaders. Whether it can hold that position is another matter — rivals are shipping new flagship models at a near-monthly pace.
