Meta's Superintelligence Lab released two new AI models on Monday, Muse Wonder and Muse Wonder Full, and this time the company is aiming squarely at the top of the leaderboard rather than the bottom of the latency chart.
The earlier Muse Spark model was built to be small and fast. Wonder and Wonder Full are bigger, Meta says, and score higher on benchmarks that measure complex work like coding. In its own testing, described in the model card, Meta puts Wonder and Wonder Full roughly level with Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra. Wired, which got an earlier look at the results, reports that Meta's own numbers place the top model just behind Claude Opus 5.5 on the Artificial Analysis Intelligence Index, an aggregated score widely used to compare model capability, while beating GPT-6 Astra and Fable 5.1 outright on some specialized tests, including ones for legal document work and software development.
That's a notable claim for a company whose last few Llama releases had researchers moving on to rivals. "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence," Meta said in its launch post.
Pricing and access
Developers reaching Muse Wonder through the Meta Model API will pay $4 per million input tokens and $20 per million output tokens, putting it in the same price bracket as GPT-6 Sol and Opus 5.5. Muse Wonder Full won't be sold through the API at all. It's reserved for subscribers to Muse Code, Meta's AI coding tool, and both models will also run inside Meta's Muse personal assistant.
Memory and focus
Meta is leaning hard on two features it calls Active Memory and Active Focus. Active Memory is meant to let the model hold onto and reuse information a user has given it over days or months rather than starting fresh every session. "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," research scientist Kavya Kopparapu said in a video accompanying the release. Meta points to MemoryBench, a benchmark testing continual learning and memory construction, where it says Wonder now outscores Fable, Opus and Astra.
Active Focus is pitched differently: it's meant to keep an AI agent locked onto a stated goal during long, open-ended tasks, the kind of job where models tend to drift or compound small errors into expensive ones. Meta also says it trained Wonder to refuse dangerous requests, such as help building bioweapons or hacking tools, and that the model scores well on benchmarks measuring resistance to that kind of prompt. The company has not said whether agents built on Wonder have misbehaved during its own testing, or what guardrails are specific to this release.
The launch is the first frontier-scale model from Meta since Scale AI founder Alexandr Wang took over its AI division in August of last year, following a reorganization that saw some researchers leave within months of joining. Meta says Wonder and Wonder Full were trained at its Prometheus data center in central Ohio, which came online earlier this year. Zuckerberg said last month on X that the release had been delayed so the team could work on safety, and Meta published a framework for scaling capability safely earlier this month. The company still has not disclosed how the Wonder models were built or trained.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.