Meta's Superintelligence Lab has released two new AI models, Muse Wonder and Muse Wonder Full, which Meta describes as its first attempt at a frontier model, the most capable class of AI systems. Meta has released AI models before, most notably its Llama family, but those were open models that anyone could download and run.

The Wonder models are closed: Meta has not released their weights and has not said whether it will.

Meta says both models score comparably to the strongest models from OpenAI and Anthropic, and better on its own memory test. Independent testing partly supports that.

Artificial Analysis, a firm that runs the same set of ten evaluations across hundreds of models, gives Muse Wonder Full a score of 56 on its Intelligence Index, tied with Anthropic's Claude Sonnet 5.5 and behind only Claude Opus 5.5 at 58. The regular Muse Wonder scores 52, roughly level with OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1, which each score 53.

The catch is access. Wonder Full, the stronger model, is only available inside Meta's own Muse Code subscription and is not offered to developers through Meta's API. Through the API, Wonder costs $4 per million input tokens and $20 per million output tokens, about R68 and R338, putting it in the same price range as the top models from OpenAI and Anthropic.

Some of Meta's headline results come from benchmarks it designed itself, including MemoryBench, which tests how well a model remembers information over long periods, and its Personal Superintelligence Index. It is also the lab's first major release since Scale AI founder Alexandr Wang took charge of Meta's AI work in August 2025, and the models were trained at Meta's new Prometheus data centre in Ohio.

TechCrunch reported the launch and Meta's focus on what it calls active memory, a model's ability to remember what matters to a user over days and months.

VentureBeat set out the Artificial Analysis scores and noted the shift away from the downloadable Llama models that made Meta popular with developers. Meta's own announcement calls the release a major leap toward personal superintelligence.

The bigger change is not the benchmark score; it is the business model. Llama made Meta the main source of capable AI that a company could download and run on its own servers, which mattered to organisations that wanted to keep data in-house, including South African firms weighing their obligations under POPIA.

Meta's best new models do not work that way. They are rented through Meta's API or used inside its own products, like the models from OpenAI and Anthropic. On performance, the honest summary is that Meta is now in the top group but not leading it, and the version developers can actually buy sits a step below the version Meta advertises. Benchmark claims deserve some caution here, because Meta was caught presenting a specially tuned version of Llama 4 on a public leaderboard last year.

Independent results like Artificial Analysis are the better guide, and the best test is still your own work. The frontier is getting crowded, and that crowding is good for buyers on price, even as labs hold back models over safety concerns.

You might also like our piece on Claude becoming a full office suite, OpenAI buying Glass Imaging, and Google testing Gemini phone calls to businesses.

Get more SA tech and business news and subscribe to The Open Letter.