Following the Llama series of models, Meta has officially released a new generation of frontier models, the Muse Wonder series, including two versions: Muse Wonder and Muse Wonder Full.
It seems Meta intends to build Muse into a new brand series. After all, it just launched the Manus-like Muse Agent, and now there is a same-named Muse series of models.
That said, I had almost forgotten that Meta is also making models. I even left Meta out of my own model price list. My apologies.
Officially, Muse Wonder is positioned as a frontier foundation for personal and professional agents. This is also the core result Meta trained in its gigawatt-scale Prometheus data center after 14 months of foundational architecture rebuilding.
For the record, this time Muse Wonder has shifted to closed source.
In terms of performance, the Full model currently matches Claude Sonnet 5.5 on the composite intelligence index, surpasses GPT-6 Astra, and the base model is second only to GPT-6 Astra.
Claude Opus 5.5 > Muse Wonder Full = Claude Sonnet 5.5 > GPT-6 Astra > Muse Wonder
This only concerns the composite intelligence index and does not represent data comparisons across vertical domains. Don't take it too seriously; it is mainly to provide a more intuitive idea.
Currently, the Full version has no public API, and the base version's API pricing matches Claude Opus 5.5. To be honest, this price is "very, very" uncompetitive. Sonnet 5.5 costs only half as much. At this price, how is it supposed to compete? It can't win standing or sitting—it can only lie down.
Right after this, Claude and GPT are about to iterate again, and Gemini 4 Argon is also about to launch. Meta's 14 months of work will immediately be ganged up on again.
A bit more technical detail: this generation of models highlights two core mechanisms: Active Memory and Active Focus.
Active Memory focuses on long-term memory and cross-session recall. It can remember a user's important background across multi-turn, long-horizon interactions and proactively retrieve and invoke it when executing specific tasks.
Active Focus emphasizes execution resilience on long-horizon tasks. When facing complex engineering, the model will keep trying alternative paths and independently troubleshoot blockers. At critical steps or when information is insufficient, it will also proactively ask the user for follow-up confirmation. Officially, it was validated on the long-horizon code benchmarks SWE-Marathon and FocusBench, and it demonstrated on site the ability to directly generate a playable 3D kitchen game from just a few images.
The official blog post also specifically addressed the phenomena of "leaderboard chasing" and "big model smell." The official point is that through extended testing, and by relying on a stronger base model, generalization ability on unseen tasks is more robust, rather than just scoring high on leaderboard-chasing datasets.
In terms of availability, Muse Wonder is available today on Meta Model API and Muse Code, while Muse Wonder Full is initially available to Muse Code subscribers, and will later be gradually integrated into everyday Muse applications and Artifact environments.