According to Wired (Oct 12, 2026) and the Latent Space AINews newsletter (Oct 13, 2026), Meta has released two new AI models, Muse Wonder and Muse Wonder Full, from its Meta Superintelligence Labs (MSL). Meta says Muse Wonder Full has 'frontier capabilities,' and the company describes the pair as 'the next step on our scaling ladder and a major leap toward personal superintelligence.' The models will be available through an API and through Muse Code, a software-writing tool, and the best version will power features including Meta's Muse personal agent. Per Wired, Meta claims the largest version outperforms Anthropic's Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index, and beats OpenAI's GPT-6 Astra on some specialized benchmarks, including legal-document work and software building. Wired notes benchmarks are not always a reliable measure of real-world performance. The models include two new features, Active Memory and Active Focus, designed to keep agents locked onto a stated goal and prevent compounding errors. Meta says it trained the model to refuse dangerous outputs such as bioweapon recipes and hacking help. Meta did not disclose how the models were trained, nor whether it observed agent misbehavior in testing. CEO Mark Zuckerberg said last month the release was delayed to work on safety.
- Two models: Muse Wonder and Muse Wonder Full, from Meta Superintelligence Labs
- Available via API and Muse Code; powers the Muse personal agent
- Meta claims it ranks just below Claude Opus 5.5 and above GPT-6 Astra on some benchmarks
- New Active Memory and Active Focus features aim to keep agents on-task
- Meta did not disclose training details or whether agents misbehaved in testing
What it means for you
Meta is back in the top tier of AI model providers, which mostly matters as competition: more frontier-grade options tends to mean better pricing and features for everyone downstream. The benchmark claims come from Meta itself and are measured on tests that don't always reflect real work, so treat 'best in class' as a marketing claim until you've tested it on your own tasks. The genuinely interesting bit is the agent-control features (Active Memory, Active Focus) aimed at stopping AI agents from drifting off-task — a real, known problem — but Meta didn't share whether those features actually prevented misbehavior in testing.
Try this
If you already run anything on a frontier model, write down the two or three tasks you actually rely on it for, then re-run those same tasks on a new model before switching. The headline benchmark means nothing; your own results do.
Who should care
Teams building or running AI agents that take multi-step actions, and anyone evaluating which model to standardize on. Especially relevant if you've hit the 'the agent lost the plot halfway through' problem.
Skip this if
You're using AI through an off-the-shelf product (a chatbot, a feature in software you already pay for) and don't choose the underlying model yourself — this changes nothing for you today.
Sources: Latent Space, Wired AI — read the original