Meta just dropped the biggest AI news of the day: Muse Wonder and Muse Wonder Full, its first true frontier models, launched Monday by Meta Superintelligence Labs. After years of playing catch-up in the model race, Meta is officially in the frontier club alongside OpenAI and Anthropic.
The details
Muse Wonder and the larger Muse Wonder Full are the pair's two flavors. Where Muse Spark was built for speed and low latency, Wonder is bigger and more capable — better benchmark scores, better complex-task performance, and better coding. Meta calls them "a major leap toward personal superintelligence."
The headline feature is what Meta calls Active Memory: the models retain and synthesize things you tell them over days, weeks, and months. In Meta's own testing, Wonder now scores higher on the MemoryBench benchmark — which tests continual learning and memory construction — than Claude Fable, Claude Opus, or GPT-6 Astra. A companion feature, Active Focus, keeps the model working through longer tasks and pushes it to try a different approach when something goes wrong.
On the independent Artificial Analysis Intelligence Index, Wonder Full scores 56 — just behind Claude Opus 5.5 (58) and ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (53 each). The regular Wonder scores 52, tying GPT-6.1 Sol. That's a giant jump from Muse Spark's 31.
Wonder is available through the Meta Model API and Muse Code at $4.00 per million input tokens and $20.00 per million output tokens — roughly the same cost tier as GPT-6 Sol or Opus 5.5. Wonder Full is only for Muse Code subscribers, not served through the API. Meta says Wonder will power its Muse personal agent starting with paying subscribers, but gave no date. Fun detail: the model was code-named "Watermelon" — Zuckerberg teased the launch with a watermelon photo on Threads.
Why it matters
Meta went from lightweight Spark models to a frontier model in about a year of Meta Superintelligence Labs. And Active Memory is the differentiator in the personal-agent race: a model that truly remembers you is the whole pitch of "personal superintelligence." The model race is now officially four-way: OpenAI, Anthropic, Google, and Meta.
The take
Two caveats. First, Meta's claims of beating GPT-6 Astra and Claude Fable 5.1 on some specialized tests come from Meta's own testing — treat them as claims, not gospel. The independent AI index score (56, right behind Opus 5.5) is the real signal: this looks genuinely competitive. Second, Meta hasn't said whether Wonder's weights will go public like the smaller Muse Glimmer — Alexandr Wang says Meta is "still committed to open source," and a full safety report is coming soon. Watch for it.
Comments
Post a Comment