Meta has launched two new AI models, Muse Wonder and Muse Wonder Full, which the company says can compete with the industry's best, according to The Verge and Wired. The models are available to Muse Code subscribers, and Wonder can also be accessed through the Meta Model API. Meta says it plans to eventually power its Muse AI agent with Wonder, starting with subscribers, though it has not provided a timeline. Citing internal testing on industry benchmarks, Meta says the largest version of the model outperforms Anthropic's Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index, and outperforms OpenAI's GPT-6 Astra on some specialized benchmarks, including legal-document and software-building tasks. Wired notes that 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 to a stated goal and prevent them from becoming confused during complex tasks. Meta says it trained Muse Wonder to refuse dangerous outputs such as bioweapon recipes or hacking assistance. The launches follow Meta's overhaul of its AI research under Meta Superintelligence Labs, led by former Scale AI CEO Alexandr Wang. Meta did not disclose how the models were built or trained.
- Muse Wonder and Muse Wonder Full are available to Muse Code subscribers; Wonder is also accessible via the Meta Model API
- Meta claims the largest model ranks just below Claude Opus 5.5 and above GPT-6 Astra on certain benchmarks, per internal testing
- New features Active Memory and Active Focus aim to keep agents on-task; Meta did not disclose whether agents misbehaved in testing
What it means for you
Another large AI lab has a new model that it says matches the leaders. For most organizations this changes nothing today — the models are aimed at developers and agent-builders, and the performance claims come from Meta's own benchmarks, which the reporting itself cautions may not reflect real work. The more interesting detail is the agent features meant to stop AI from drifting off-task, which is a real and expensive problem when you let software run long, open-ended jobs.
Try this
If any team is running or piloting AI agents on multi-step work, write down exactly where you've seen one wander off-goal or compound a mistake — that record tells you whether features like these would actually help you, far better than a vendor benchmark.
Who should care
Developers and technical teams building or evaluating AI agents, and whoever owns vendor selection for AI tooling.
Skip this if
You're not building with AI APIs or agents — this is model-race positioning with no bearing on your day-to-day operations.
Sources: The Verge AI, Wired AI — read the original